project_portfolio/03-literature-review.html at main · agodinezmm2007/project_portfolio · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html xmlns="http://www.w3.org/1999/xhtml" lang="en" xml:lang="en"><head>

<meta charset="utf-8">
<meta name="generator" content="quarto-1.5.57">

<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes">


<title>4&nbsp; Literature Review – Portfolio</title>
<style>
code{white-space: pre-wrap;}
span.smallcaps{font-variant: small-caps;}
div.columns{display: flex; gap: min(4vw, 1.5em);}
div.column{flex: auto; overflow-x: auto;}
div.hanging-indent{margin-left: 1.5em; text-indent: -1.5em;}
ul.task-list{list-style: none;}
ul.task-list li input[type="checkbox"] {
  width: 0.8em;
  margin: 0 0.8em 0.2em -1em; /* quarto-specific, see https://github.com/quarto-dev/quarto-cli/issues/4556 */
  vertical-align: middle;
}
/* CSS for syntax highlighting */
pre > code.sourceCode { white-space: pre; position: relative; }
pre > code.sourceCode > span { line-height: 1.25; }
pre > code.sourceCode > span:empty { height: 1.2em; }
.sourceCode { overflow: visible; }
code.sourceCode > span { color: inherit; text-decoration: inherit; }
div.sourceCode { margin: 1em 0; }
pre.sourceCode { margin: 0; }
@media screen {
div.sourceCode { overflow: auto; }
}
@media print {
pre > code.sourceCode { white-space: pre-wrap; }
pre > code.sourceCode > span { display: inline-block; text-indent: -5em; padding-left: 5em; }
}
pre.numberSource code
  { counter-reset: source-line 0; }
pre.numberSource code > span
  { position: relative; left: -4em; counter-increment: source-line; }
pre.numberSource code > span > a:first-child::before
  { content: counter(source-line);
    position: relative; left: -1em; text-align: right; vertical-align: baseline;
    border: none; display: inline-block;
    -webkit-touch-callout: none; -webkit-user-select: none;
    -khtml-user-select: none; -moz-user-select: none;
    -ms-user-select: none; user-select: none;
    padding: 0 4px; width: 4em;
  }
pre.numberSource { margin-left: 3em;  padding-left: 4px; }
div.sourceCode
  {   }
@media screen {
pre > code.sourceCode > span > a:first-child::before { text-decoration: underline; }
}
/* CSS for citations */
div.csl-bib-body { }
div.csl-entry {
  clear: both;
  margin-bottom: 0em;
}
.hanging-indent div.csl-entry {
  margin-left:2em;
  text-indent:-2em;
}
div.csl-left-margin {
  min-width:2em;
  float:left;
}
div.csl-right-inline {
  margin-left:2em;
  padding-left:1em;
}
div.csl-indent {
  margin-left: 2em;
}</style>


<script src="site_libs/quarto-nav/quarto-nav.js"></script>
<script src="site_libs/quarto-nav/headroom.min.js"></script>
<script src="site_libs/clipboard/clipboard.min.js"></script>
<script src="site_libs/quarto-search/autocomplete.umd.js"></script>
<script src="site_libs/quarto-search/fuse.min.js"></script>
<script src="site_libs/quarto-search/quarto-search.js"></script>
<meta name="quarto:offset" content="./">
<link href="./04-prelim-analysis.html" rel="next">
<link href="./02-data-collection.html" rel="prev">
<link href="./pic/cover.png" rel="icon" type="image/png">
<script src="site_libs/quarto-html/quarto.js"></script>
<script src="site_libs/quarto-html/popper.min.js"></script>
<script src="site_libs/quarto-html/tippy.umd.min.js"></script>
<script src="site_libs/quarto-html/anchor.min.js"></script>
<link href="site_libs/quarto-html/tippy.css" rel="stylesheet">
<link href="site_libs/quarto-html/quarto-syntax-highlighting.css" rel="stylesheet" id="quarto-text-highlighting-styles">
<script src="site_libs/bootstrap/bootstrap.min.js"></script>
<link href="site_libs/bootstrap/bootstrap-icons.css" rel="stylesheet">
<link href="site_libs/bootstrap/bootstrap.min.css" rel="stylesheet" id="quarto-bootstrap" data-mode="light">
<script id="quarto-search-options" type="application/json">{
  "location": "sidebar",
  "copy-button": false,
  "collapse-after": 3,
  "panel-placement": "start",
  "type": "textbox",
  "limit": 50,
  "keyboard-shortcut": [
    "f",
    "/",
    "s"
  ],
  "show-item-context": false,
  "language": {
    "search-no-results-text": "No results",
    "search-matching-documents-text": "matching documents",
    "search-copy-link-title": "Copy link to search",
    "search-hide-matches-text": "Hide additional matches",
    "search-more-match-text": "more match in this document",
    "search-more-matches-text": "more matches in this document",
    "search-clear-button-title": "Clear",
    "search-text-placeholder": "",
    "search-detached-cancel-button-title": "Cancel",
    "search-submit-button-title": "Submit",
    "search-label": "Search"
  }
}</script>
<style>html{ scroll-behavior: smooth; }</style>


</head>

<body class="nav-sidebar docked">

<div id="quarto-search-results"></div>
  <header id="quarto-header" class="headroom fixed-top">
  <nav class="quarto-secondary-nav">
    <div class="container-fluid d-flex">
      <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
        <i class="bi bi-layout-text-sidebar-reverse"></i>
      </button>
        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="./03-literature-review.html"><span class="chapter-number">4</span>&nbsp; <span class="chapter-title">Literature Review</span></a></li></ol></nav>
        <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
        </a>
      <button type="button" class="btn quarto-search-button" aria-label="Search" onclick="window.quartoOpenSearch();">
        <i class="bi bi-search"></i>
      </button>
    </div>
  </nav>
</header>
<!-- content -->
<div id="quarto-content" class="quarto-container page-columns page-rows-contents page-layout-article">
<!-- sidebar -->
  <nav id="quarto-sidebar" class="sidebar collapse collapse-horizontal quarto-sidebar-collapse-item sidebar-navigation docked overflow-auto">
    <div class="pt-lg-2 mt-2 text-left sidebar-header">
    <div class="sidebar-title mb-0 py-0">
      <a href="./">Portfolio</a>
        <div class="sidebar-tools-main">
  <a href="" class="quarto-reader-toggle quarto-navigation-tool px-1" onclick="window.quartoToggleReader(); return false;" title="Toggle reader mode">
  <div class="quarto-reader-toggle-btn">
  <i class="bi"></i>
  </div>
</a>
</div>
    </div>
      </div>
        <div class="mt-2 flex-shrink-0 align-items-center">
        <div class="sidebar-search">
        <div id="quarto-search" class="" title="Search"></div>
        </div>
        </div>
    <div class="sidebar-menu-container">
    <ul class="list-unstyled mt-1">
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./index.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text">Preface</span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./intro.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">1</span>&nbsp; <span class="chapter-title">Introduction</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./01-proposal.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">2</span>&nbsp; <span class="chapter-title">Literature Analytics and Methodological Gap Analysis in Geospatial Epidemiology</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./02-data-collection.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">3</span>&nbsp; <span class="chapter-title">Data Collection</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./03-literature-review.html" class="sidebar-item-text sidebar-link active">
 <span class="menu-text"><span class="chapter-number">4</span>&nbsp; <span class="chapter-title">Literature Review</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./04-prelim-analysis.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">5</span>&nbsp; <span class="chapter-title">Preliminary Analysis</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./05-aim3.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">6</span>&nbsp; <span class="chapter-title">Proposed Deep Learning Framework for UOG Site Monitoring (Aim 3)</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./06-technical-report.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">7</span>&nbsp; <span class="chapter-title">Technical Report</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./summary.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">8</span>&nbsp; <span class="chapter-title">Summary</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./references.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">9</span>&nbsp; <span class="chapter-title">References</span></span></a>
  </div>
</li>
        <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./update.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text">Project Update</span></a>
  </div>
</li>
        <li class="sidebar-item sidebar-item-section">
      <div class="sidebar-item-container">
            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-1" role="navigation" aria-expanded="true">
 <span class="menu-text">Appendices</span></a>
          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-1" role="navigation" aria-expanded="true" aria-label="Toggle section">
            <i class="bi bi-chevron-right ms-2"></i>
          </a>
      </div>
      <ul id="quarto-sidebar-section-1" class="collapse list-unstyled sidebar-section depth1 show">
          <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./appendix-a-etl-rag.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">A</span>&nbsp; <span class="chapter-title">appendix-a-etl-rag.html</span></span></a>
  </div>
</li>
          <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./appendix-b-lca-rf.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">B</span>&nbsp; <span class="chapter-title">Preliminary LCA Cover crop Analysis</span></span></a>
  </div>
</li>
          <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./appendix-c-hardware.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">C</span>&nbsp; <span class="chapter-title">Appendix C: Computational Hardware</span></span></a>
  </div>
</li>
          <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./appendix-d-neo4j-queries.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">D</span>&nbsp; <span class="chapter-title">Knowledge Graph Queries and Supporting Data</span></span></a>
  </div>
</li>
          <li class="sidebar-item">
  <div class="sidebar-item-container">
  <a href="./appendix-e-db-queries.html" class="sidebar-item-text sidebar-link">
 <span class="menu-text"><span class="chapter-number">E</span>&nbsp; <span class="chapter-title">Database Queries</span></span></a>
  </div>
</li>
      </ul>
  </li>
    </ul>
    </div>
</nav>
<div id="quarto-sidebar-glass" class="quarto-sidebar-collapse-item" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item"></div>
<!-- margin-sidebar -->
    <div id="quarto-margin-sidebar" class="sidebar margin-sidebar">
        <nav id="TOC" role="doc-toc" class="toc-active">
    <h2 id="toc-title">Table of contents</h2>

  <ul>
  <li><a href="#introduction" id="toc-introduction" class="nav-link active" data-scroll-target="#introduction"><span class="header-section-number">4.1</span> Introduction:</a></li>
  <li><a href="#sec-methods" id="toc-sec-methods" class="nav-link" data-scroll-target="#sec-methods"><span class="header-section-number">4.2</span> Methods</a>
  <ul>
  <li><a href="#eligibility-criteria" id="toc-eligibility-criteria" class="nav-link" data-scroll-target="#eligibility-criteria"><span class="header-section-number">4.2.1</span> Eligibility Criteria</a></li>
  <li><a href="#sec-methods-search" id="toc-sec-methods-search" class="nav-link" data-scroll-target="#sec-methods-search"><span class="header-section-number">4.2.2</span> Information Sources and Search Strategy</a></li>
  <li><a href="#sec-methods-selection" id="toc-sec-methods-selection" class="nav-link" data-scroll-target="#sec-methods-selection"><span class="header-section-number">4.2.3</span> Study Selection and Data Management</a>
  <ul class="collapse">
  <li><a href="#initial-record-collection-and-pdf-retrieval" id="toc-initial-record-collection-and-pdf-retrieval" class="nav-link" data-scroll-target="#initial-record-collection-and-pdf-retrieval"><span class="header-section-number">4.2.3.1</span> Initial Record Collection and PDF Retrieval</a></li>
  <li><a href="#data-extraction-and-knowledge-graph-preprocessing" id="toc-data-extraction-and-knowledge-graph-preprocessing" class="nav-link" data-scroll-target="#data-extraction-and-knowledge-graph-preprocessing"><span class="header-section-number">4.2.3.2</span> Data Extraction and Knowledge Graph Preprocessing</a></li>
  <li><a href="#knowledge-graph-construction" id="toc-knowledge-graph-construction" class="nav-link" data-scroll-target="#knowledge-graph-construction"><span class="header-section-number">4.2.3.3</span> Knowledge Graph Construction</a></li>
  <li><a href="#sec-methods-eligibility" id="toc-sec-methods-eligibility" class="nav-link" data-scroll-target="#sec-methods-eligibility"><span class="header-section-number">4.2.3.4</span> Eligibility Assessment using the Knowledge Graph</a></li>
  </ul></li>
  <li><a href="#sec-methods-final" id="toc-sec-methods-final" class="nav-link" data-scroll-target="#sec-methods-final"><span class="header-section-number">4.2.4</span> Final Included Studies and Data Synthesis</a></li>
  </ul></li>
  <li><a href="#health-impacts-of-ozone" id="toc-health-impacts-of-ozone" class="nav-link" data-scroll-target="#health-impacts-of-ozone"><span class="header-section-number">4.3</span> Health Impacts of Ozone:</a></li>
  <li><a href="#geospatial-epidemiological-methods" id="toc-geospatial-epidemiological-methods" class="nav-link" data-scroll-target="#geospatial-epidemiological-methods"><span class="header-section-number">4.4</span> Geospatial Epidemiological Methods:</a></li>
  <li><a href="#machine-learning-in-ozone-epidemiology" id="toc-machine-learning-in-ozone-epidemiology" class="nav-link" data-scroll-target="#machine-learning-in-ozone-epidemiology"><span class="header-section-number">4.5</span> Machine Learning in Ozone Epidemiology:</a>
  <ul>
  <li><a href="#frequently-applied-machine-learning-techniques" id="toc-frequently-applied-machine-learning-techniques" class="nav-link" data-scroll-target="#frequently-applied-machine-learning-techniques"><span class="header-section-number">4.5.1</span> Frequently Applied Machine Learning Techniques</a></li>
  <li><a href="#limitations-of-conventional-machine-learning-methods" id="toc-limitations-of-conventional-machine-learning-methods" class="nav-link" data-scroll-target="#limitations-of-conventional-machine-learning-methods"><span class="header-section-number">4.5.2</span> Limitations of Conventional Machine Learning Methods</a></li>
  <li><a href="#applications-of-cnn-and-gnn-in-ozone-and-cardiovascular-epidemiology" id="toc-applications-of-cnn-and-gnn-in-ozone-and-cardiovascular-epidemiology" class="nav-link" data-scroll-target="#applications-of-cnn-and-gnn-in-ozone-and-cardiovascular-epidemiology"><span class="header-section-number">4.5.3</span> Applications of CNN and GNN in Ozone and Cardiovascular Epidemiology</a></li>
  </ul></li>
  <li><a href="#summary-of-gaps" id="toc-summary-of-gaps" class="nav-link" data-scroll-target="#summary-of-gaps"><span class="header-section-number">4.6</span> Summary of Gaps:</a>
  <ul>
  <li><a href="#current-gaps-in-cnngnn-application-for-ozone-cardiovascular-health-studies" id="toc-current-gaps-in-cnngnn-application-for-ozone-cardiovascular-health-studies" class="nav-link" data-scroll-target="#current-gaps-in-cnngnn-application-for-ozone-cardiovascular-health-studies"><span class="header-section-number">4.6.1</span> Current Gaps in CNN/GNN Application for Ozone-Cardiovascular Health Studies</a></li>
  <li><a href="#calls-for-advanced-cnngnn-spatial-temporal-methods" id="toc-calls-for-advanced-cnngnn-spatial-temporal-methods" class="nav-link" data-scroll-target="#calls-for-advanced-cnngnn-spatial-temporal-methods"><span class="header-section-number">4.6.2</span> Calls for Advanced CNN/GNN Spatial-Temporal Methods</a></li>
  </ul></li>
  <li><a href="#graph-neural-networks-gnns" id="toc-graph-neural-networks-gnns" class="nav-link" data-scroll-target="#graph-neural-networks-gnns"><span class="header-section-number">4.7</span> Graph Neural Networks (GNNs)</a>
  <ul>
  <li><a href="#recommended-gnn-architectures-for-spatiotemporal-ozone-modeling" id="toc-recommended-gnn-architectures-for-spatiotemporal-ozone-modeling" class="nav-link" data-scroll-target="#recommended-gnn-architectures-for-spatiotemporal-ozone-modeling"><span class="header-section-number">4.7.1</span> Recommended GNN Architectures for Spatiotemporal Ozone Modeling</a></li>
  <li><a href="#relevance-of-gnn-approaches-to-cardiovascular-epidemiology" id="toc-relevance-of-gnn-approaches-to-cardiovascular-epidemiology" class="nav-link" data-scroll-target="#relevance-of-gnn-approaches-to-cardiovascular-epidemiology"><span class="header-section-number">4.7.2</span> Relevance of GNN Approaches to Cardiovascular Epidemiology</a></li>
  </ul></li>
  <li><a href="#convolutional-neural-networks-cnns" id="toc-convolutional-neural-networks-cnns" class="nav-link" data-scroll-target="#convolutional-neural-networks-cnns"><span class="header-section-number">4.8</span> Convolutional Neural Networks (CNNs)</a>
  <ul>
  <li><a href="#previous-applications-of-cnns-in-air-pollution-exposure-modeling" id="toc-previous-applications-of-cnns-in-air-pollution-exposure-modeling" class="nav-link" data-scroll-target="#previous-applications-of-cnns-in-air-pollution-exposure-modeling"><span class="header-section-number">4.8.1</span> Previous Applications of CNNs in Air Pollution Exposure Modeling</a></li>
  <li><a href="#cnn-applications-directly-linking-air-pollution-to-cardiovascular-outcomes" id="toc-cnn-applications-directly-linking-air-pollution-to-cardiovascular-outcomes" class="nav-link" data-scroll-target="#cnn-applications-directly-linking-air-pollution-to-cardiovascular-outcomes"><span class="header-section-number">4.8.2</span> CNN Applications Directly Linking Air Pollution to Cardiovascular Outcomes</a></li>
  </ul></li>
  <li><a href="#identifying-high-risk-areas" id="toc-identifying-high-risk-areas" class="nav-link" data-scroll-target="#identifying-high-risk-areas"><span class="header-section-number">4.9</span> Identifying High-Risk Areas</a>
  <ul>
  <li><a href="#statistical-methods-and-models-for-area-level-risk-estimation" id="toc-statistical-methods-and-models-for-area-level-risk-estimation" class="nav-link" data-scroll-target="#statistical-methods-and-models-for-area-level-risk-estimation"><span class="header-section-number">4.9.1</span> Statistical Methods and Models for Area-Level Risk Estimation</a></li>
  <li><a href="#spatial-analytical-approaches-for-high-risk-area-identification" id="toc-spatial-analytical-approaches-for-high-risk-area-identification" class="nav-link" data-scroll-target="#spatial-analytical-approaches-for-high-risk-area-identification"><span class="header-section-number">4.9.2</span> Spatial Analytical Approaches for High-Risk Area Identification</a></li>
  </ul></li>
  <li><a href="#quantifying-ozone-impact" id="toc-quantifying-ozone-impact" class="nav-link" data-scroll-target="#quantifying-ozone-impact"><span class="header-section-number">4.10</span> Quantifying Ozone Impact</a>
  <ul>
  <li><a href="#empirical-estimates-from-short-term-exposure" id="toc-empirical-estimates-from-short-term-exposure" class="nav-link" data-scroll-target="#empirical-estimates-from-short-term-exposure"><span class="header-section-number">4.10.1</span> Empirical Estimates from Short-Term Exposure</a></li>
  </ul></li>
  <li><a href="#empirical-estimates-from-mortality-or-long-term-exposure" id="toc-empirical-estimates-from-mortality-or-long-term-exposure" class="nav-link" data-scroll-target="#empirical-estimates-from-mortality-or-long-term-exposure"><span class="header-section-number">4.11</span> Empirical Estimates from Mortality or Long-Term Exposure</a></li>
  <li><a href="#exploring-non-linearity" id="toc-exploring-non-linearity" class="nav-link" data-scroll-target="#exploring-non-linearity"><span class="header-section-number">4.12</span> Exploring Non-Linearity</a>
  <ul>
  <li><a href="#studies-reporting-non-linear-relationships" id="toc-studies-reporting-non-linear-relationships" class="nav-link" data-scroll-target="#studies-reporting-non-linear-relationships"><span class="header-section-number">4.12.1</span> Studies Reporting Non-linear Relationships</a></li>
  <li><a href="#studies-reporting-no-observed-non-linear-relationships" id="toc-studies-reporting-no-observed-non-linear-relationships" class="nav-link" data-scroll-target="#studies-reporting-no-observed-non-linear-relationships"><span class="header-section-number">4.12.2</span> Studies Reporting No Observed Non-linear Relationships</a></li>
  </ul></li>
  <li><a href="#spatial-spillover" id="toc-spatial-spillover" class="nav-link" data-scroll-target="#spatial-spillover"><span class="header-section-number">4.13</span> Spatial Spillover</a></li>
  <li><a href="#validation-and-predictive-accuracy" id="toc-validation-and-predictive-accuracy" class="nav-link" data-scroll-target="#validation-and-predictive-accuracy"><span class="header-section-number">4.14</span> Validation and Predictive Accuracy</a>
  <ul>
  <li><a href="#validation-metrics-for-cnn-and-gnn-models" id="toc-validation-metrics-for-cnn-and-gnn-models" class="nav-link" data-scroll-target="#validation-metrics-for-cnn-and-gnn-models"><span class="header-section-number">4.14.1</span> Validation Metrics for CNN and GNN Models</a></li>
  <li><a href="#predictive-accuracy-of-random-forest-models" id="toc-predictive-accuracy-of-random-forest-models" class="nav-link" data-scroll-target="#predictive-accuracy-of-random-forest-models"><span class="header-section-number">4.14.2</span> Predictive Accuracy of Random Forest Models</a></li>
  </ul></li>
  <li><a href="#discussion" id="toc-discussion" class="nav-link" data-scroll-target="#discussion"><span class="header-section-number">4.15</span> Discussion</a></li>
  <li><a href="#future-work" id="toc-future-work" class="nav-link" data-scroll-target="#future-work"><span class="header-section-number">4.16</span> Future Work</a></li>
  <li><a href="#conclusion" id="toc-conclusion" class="nav-link" data-scroll-target="#conclusion"><span class="header-section-number">4.17</span> Conclusion</a></li>
  <li><a href="#references" id="toc-references" class="nav-link" data-scroll-target="#references"><span class="header-section-number">5</span> References</a></li>
  </ul>
</nav>
    </div>
<!-- main -->
<main class="content" id="quarto-document-content">

<header id="title-block-header" class="quarto-title-block default">
<div class="quarto-title">
<div class="quarto-title-block"><div><h1 class="title"><span class="chapter-number">4</span>&nbsp; <span class="chapter-title">Literature Review</span></h1><button type="button" class="btn code-tools-button" id="quarto-code-tools-source"><i class="bi"></i> Code</button></div></div>
</div>


<div class="quarto-title-meta">


  </div>


</header>


<p>#03-literature-review.qmd</p>
<section id="introduction" class="level2" data-number="4.1">
<h2 data-number="4.1" class="anchored" data-anchor-id="introduction"><span class="header-section-number">4.1</span> Introduction:</h2>
<p>The rapidly expanding volume of scientific literature presents considerable challenges for researchers, particularly in complex, interdisciplinary fields such as geospatial epidemiology, where manually synthesizing vast amounts of information to identify emergent methodological trends and pinpoint specific research gaps can be exceptionally demanding and time-consuming. To address these limitations, an automated pipeline leveraging Extract, Transform, Load (ETL) processes and Retrieval-Augmented Generation (RAG) has been developed as outlined in our broader project aimed at enhancing large-scale academic literature analytics and methodological gap analysis. This system is designed to systematically process extensive collections of academic articles, extract structured text and key information including domain-specific metrics and methodologies, and thereby facilitate the efficient identification of under-explored research areas and novel or underutilized analytical techniques.</p>
<p>This literature review serves as a practical application and demonstration of the aforementioned AI-powered pipeline’s capabilities in navigating a complex research domain. The chosen area for this demonstration is the intersection of ambient ozone exposure, cardiovascular (specifically heart) disease outcomes, and the application of geospatial and machine learning methodologies, with a particular emphasis on deep learning techniques such as Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs). This thematic focus was strategically selected not only due to the pressing public health questions involved but also facilitated by prior access to extensive, highly granular datasets, including daily census-tract level ozone concentrations and CDC cardiovascular disease mortality data stratified by age group, sex, and race. Initial explorations using the pipeline further sharpened this focus by highlighting a notable underutilization of sophisticated neural network architectures like CNNs and GNNs in studies specifically linking ozone exposure to cardiovascular disease, despite their considerable potential for modeling intricate spatiotemporal relationships.</p>
<p>Therefore, the objectives of this review are: (1) to present a comprehensive overview of the current research landscape concerning ozone, heart disease, and the use of geospatial and machine learning methods, particularly CNNs and GNNs, as identified and synthesized through our automated pipeline; (2) to detail the methodological trends, common practices, and critically, the specific research and methodological gaps revealed by this AI-assisted analysis, such as the limited application of deep learning or the scarcity of studies focusing exclusively on ozone and cardiovascular disease; and (3) to thereby provide an evidence-based foundation that will directly inform and guide subsequent empirical research, specifically the planned application of novel or underutilized analytical techniques to address these identified gaps.</p>
</section>
<section id="sec-methods" class="level2" data-number="4.2">
<h2 data-number="4.2" class="anchored" data-anchor-id="sec-methods"><span class="header-section-number">4.2</span> Methods</h2>
<p>This literature review employed a semi-automated pipeline, leveraging a custom-built Extract, Transform, Load (ETL) process and a Neo4j knowledge graph, to identify, screen, and categorize relevant literature. The entire workflow, from data acquisition to final study selection and characterization, was designed to be transparent and reproducible. The study selection process adapted the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines <a href="#fig-prisma" class="quarto-xref">Figure&nbsp;<span>4.1</span></a>.</p>
<div id="fig-prisma" class="quarto-float quarto-figure quarto-figure-left anchored">
<figure class="quarto-float quarto-float-fig figure">
<div aria-describedby="fig-prisma-caption-0ceaefa1-69ba-4598-a22c-09a6ac19f8ca">
<img src="figures/diagram2.PNG" class="img-fluid figure-img">
</div>
<figcaption class="quarto-float-caption-bottom quarto-float-caption quarto-float-fig" id="fig-prisma-caption-0ceaefa1-69ba-4598-a22c-09a6ac19f8ca">
Figure&nbsp;4.1: PRISMA Flow Diagram illustrating the literature selection process. Reasons for exclusion: Reason 1: Missing study type or study type null (n=145); Reason 2: Non-empirical study (e.g., review, meta-analysis, opinion, report) (n=60); Reason 3: Empirical study not using machine learning methods (n=159); Reason 4: Empirical ML study not focused on both ozone AND heart disease (n=39). 1260 records identified, 821 not retrieved, 439 assessed for eligibility, 403 excluded, and 36 studies included, with a further 2 manually excluded, resulting in 34 final studies.
</figcaption>
</figure>
</div>
<section id="eligibility-criteria" class="level3" data-number="4.2.1">
<h3 data-number="4.2.1" class="anchored" data-anchor-id="eligibility-criteria"><span class="header-section-number">4.2.1</span> Eligibility Criteria</h3>
<p>This review focused on studies investigating the intersection of ambient ozone exposure, cardiovascular (specifically heart) disease outcomes, and the application of geospatial and/or machine learning methodologies. To be included, studies had to: 1. Be empirical research (not reviews, commentaries, meta-analyses, reports, dissertations, etc.). 2. Explicitly identify ozone as a pollutant of interest. 3. Report an association with or investigation of heart disease or broader cardiovascular disease outcomes. 4. Employ one or more machine learning techniques in their analysis or modeling. 5. Be published as full-text articles. Studies focusing exclusively on non-ozone pollutants, non-cardiovascular outcomes, or those not utilizing machine learning were excluded. Commentaries and replies to commentaries were also excluded during a final review stage.</p>
</section>
<section id="sec-methods-search" class="level3" data-number="4.2.2">
<h3 data-number="4.2.2" class="anchored" data-anchor-id="sec-methods-search"><span class="header-section-number">4.2.2</span> Information Sources and Search Strategy</h3>
<p>A literature search was conducted across three primary scholarly databases: Scopus (via Elsevier API), OpenAlex, and PubMed. To ensure comprehensive coverage of the multidisciplinary research area, over 30 distinct search query combinations were developed (See Appendix E <a href="appendix-e-db-queries.html" class="quarto-xref"><span>Appendix E</span></a>). These queries combined keywords and concepts related to “ozone exposure,” “heart disease,” “cardiovascular disease,” “geospatial epidemiology,” “spatiotemporal analysis,” “machine learning,” and various specific geospatial and ML techniques (e.g., “GIS,” “Random Forest,” “neural networks”). The complete list of search query terms is provided in Supplementary Appendix E.</p>
<p>The execution of these searches was automated using custom Python modules (<code>fast_pubmed.py</code>, <code>fast_openalex.py</code>, <code>etl_elsevier.py</code>), which facilitated efficient querying and metadata retrieval from each database API.</p>
</section>
<section id="sec-methods-selection" class="level3" data-number="4.2.3">
<h3 data-number="4.2.3" class="anchored" data-anchor-id="sec-methods-selection"><span class="header-section-number">4.2.3</span> Study Selection and Data Management</h3>
<section id="initial-record-collection-and-pdf-retrieval" class="level4" data-number="4.2.3.1">
<h4 data-number="4.2.3.1" class="anchored" data-anchor-id="initial-record-collection-and-pdf-retrieval"><span class="header-section-number">4.2.3.1</span> Initial Record Collection and PDF Retrieval</h4>
<p>The combined searches initially yielded <strong>1,260 records</strong> (see <a href="#fig-prisma" class="quarto-xref">Figure&nbsp;<span>4.1</span></a>). Metadata for these records, including DOI, title, authors, retraction status, citation counts, abstract (where available), and publication details, were programmatically collected.</p>
<p>Full-text PDF acquisition was then attempted for all identified records. This was automated using available DOIs to query Unpaywall (via <code>async_unpaywall.py</code>) and by extracting direct PDF links from Crossref metadata (obtained through the Elsevier and OpenAlex data fetching modules). This process successfully retrieved <strong>439 unique full-text PDF articles</strong>. Records for which full-text PDFs could not be accessed (n=821) were excluded at this stage.</p>
</section>
<section id="data-extraction-and-knowledge-graph-preprocessing" class="level4" data-number="4.2.3.2">
<h4 data-number="4.2.3.2" class="anchored" data-anchor-id="data-extraction-and-knowledge-graph-preprocessing"><span class="header-section-number">4.2.3.2</span> Data Extraction and Knowledge Graph Preprocessing</h4>
<p>The 439 retrieved PDF articles were processed through a custom ETL pipeline to extract and structure relevant information:</p>
<ol type="1">
<li><p><strong>Text and Content Extraction:</strong> Full textual content, along with structured elements such as tables and LaTeX formulas, was extracted using a multiprocessing-optimized pipeline leveraging Docling’s machine learning layout analysis capabilities (<code>docling_extract_formulas_mp_multi.py</code>).</p></li>
<li><p><strong>LLM-based Field Extraction:</strong> Key metadata fields, study characteristics (e.g., study type, pollutant terms, heart diseases mentioned, ML methods used), and other relevant entities were identified and extracted from the text using a Large Language Model (LLM)-based pipeline (<code>field_extraction.py</code>). This process included robust JSON parsing, response caching, and dynamic text chunking.</p></li>
<li><p><strong>Terminology Unification:</strong> Extracted terms across various fields were standardized using a comprehensive set of custom synonym dictionaries (managed via <code>dictionaries_gui.py</code>), employing semantic similarity against precomputed embeddings to map variant terms to canonical labels. This step was crucial for ensuring consistency in the subsequent knowledge graph.</p></li>
<li><p><strong>Topic Modeling:</strong> To further characterize the thematic landscape of the 439 retrieved full-text articles, topic modeling was performed using Latent Dirichlet Allocation (LDA) facilitated by our <code>topic_modeling_culda.py</code> module. This process involved robust text preprocessing (including tokenization, stopword removal, and lemmatization) and n-gram detection. Hyperparameter tuning for the LDA model was conducted to optimize topic coherence, resulting in a final model with 26 topics and a coherence score (CV) of approximately 0.61. The specific optimized hyperparameters included settings for the number of passes, iterations, n-gram thresholds, and document frequency cutoffs (best params: <code>{'num_topics': 26, 'passes': 65, 'iterations': 200, 'bigram_min_count': 3, 'bigram_threshold': 84, 'trigram_threshold': 140, 'no_below': 7, 'no_above': 0.5}</code>). This procedure successfully generated interpretable thematic labels for the identified topics (e.g., ‘Air Quality Forecasting Models’, ‘Satellite Aerosol Monitoring Uncertainty’, ‘Geospatial Health Risk Mapping’), providing an additional layer of content characterization for the broader corpus prior to final eligibility assessment.</p></li>
<li><p><strong>Data Structuring:</strong> All extracted and processed information was compiled into structured Feather files.</p></li>
</ol>
</section>
<section id="knowledge-graph-construction" class="level4" data-number="4.2.3.3">
<h4 data-number="4.2.3.3" class="anchored" data-anchor-id="knowledge-graph-construction"><span class="header-section-number">4.2.3.3</span> Knowledge Graph Construction</h4>
<p>The structured data from the 439 articles was integrated into a Neo4j knowledge graph using a dedicated Python script (<code>kg_pipeline_gui.py</code>). This script created nodes for articles (identified by DOI) and linked them to nodes representing entities such as Study Types, Pollutant Terms, Heart Diseases, ML Methods, and Analytical Tools, based on the extracted and unified data. Relationships between these entities were explicitly modeled (e.g., <code>(Article)-[:STUDY_TYPE]-&gt;(StudyType)</code>). Uniqueness constraints were applied to key node properties (e.g., <code>Article.doi</code>, <code>Author.canonicalName</code>) to ensure data integrity.</p>
</section>
<section id="sec-methods-eligibility" class="level4" data-number="4.2.3.4">
<h4 data-number="4.2.3.4" class="anchored" data-anchor-id="sec-methods-eligibility"><span class="header-section-number">4.2.3.4</span> Eligibility Assessment using the Knowledge Graph</h4>
<p>The Neo4j knowledge graph served as the primary tool for the systematic application of eligibility criteria to the 439 full-text articles. A series of Cypher queries were developed to categorize articles and identify those meeting the review’s inclusion criteria. This knowledge graph-driven approach allowed for precise and reproducible filtering.</p>
<p>The exclusion criteria, applied sequentially based on the comprehensive categorization query (see Appendix D, Section <a href="appendix-d-neo4j-queries.html#sec-appendix-d-neo4j-study-select" class="quarto-xref"><span>Section D.3</span></a>), were as follows:</p>
<ol type="1">
<li><p><strong>Articles with missing or null study type information (n=145):</strong> Excluded if a study type could not be robustly determined or was absent in the knowledge graph.</p></li>
<li><p><strong>Non-empirical studies (n=60):</strong> Articles classified with study types such as ‘review’, ‘systematic review’, ‘meta-analysis’, ‘expert opinion’, ‘report’, ‘dissertation/thesis’, ‘short communication’, ‘methodological paper’, or ‘theoretical study’. Articles were prioritized into this category if any of their identified study types matched this list.</p></li>
<li><p><strong>Empirical studies not using Machine Learning methods (n=159):</strong> Articles identified as empirical but lacking an association with any ML Method node in the knowledge graph.</p></li>
<li><p><strong>Empirical ML studies not focused on both ozone AND heart disease (n=39):</strong> Articles that were empirical and used ML but did not have established links to both ‘ozone’ (as a PollutantTerm) AND a relevant HeartDisease node.</p></li>
</ol>
<p>This graph-based filtering process initially yielded 36 potentially eligible studies.</p>
</section>
</section>
<section id="sec-methods-final" class="level3" data-number="4.2.4">
<h3 data-number="4.2.4" class="anchored" data-anchor-id="sec-methods-final"><span class="header-section-number">4.2.4</span> Final Included Studies and Data Synthesis</h3>
<p>The 36 studies identified through the knowledge graph queries underwent a final manual review of their titles and abstracts to ensure they represented primary research contributions. At this stage, <strong>two articles were excluded</strong> as they were determined to be a commentary and a reply to a commentary, respectively.</p>
<p>This resulted in a final set of <strong>34 unique studies</strong> being included in the literature review. Data on study characteristics, pollutants, machine learning methods, heart diseases, and other analytical tools for these 34 studies were extracted from the knowledge graph (see Appendix D, Section <a href="appendix-d-neo4j-queries.html#sec-appendix-d-included-studies-char" class="quarto-xref"><span>Section D.4</span></a> for detailed distributions and queries). The synthesis of findings from these studies focused on identifying methodological gaps and common practices in the application of geospatial and machine learning techniques to ozone and heart disease research. Bibliographic data for all processed articles, including the final included set, were managed using Zotero, facilitated by programmatic integration (<code>fast_zotero.py</code>).</p>
<hr>
</section>
</section>
<section id="health-impacts-of-ozone" class="level2" data-number="4.3">
<h2 data-number="4.3" class="anchored" data-anchor-id="health-impacts-of-ozone"><span class="header-section-number">4.3</span> Health Impacts of Ozone:</h2>
<p>Empirical evidence indicates that ozone exposure affects the cardiovascular system. Short-term increases in ambient O₃ have been associated with modest but statistically significant alterations in cardiovascular physiological markers, such as decreases in systolic and diastolic blood pressure in middle-aged and older adults <span class="citation" data-cites="tang_association_2024">(<a href="#ref-tang_association_2024" role="doc-biblioref">Tang et al. 2024</a>)</span>. At the population level, interventions targeting traffic emissions that resulted in a 5.8% reduction in ambient ozone concentrations were directly correlated with a 5.9% decline in cardiovascular mortality, suggesting that lowering ozone exposure can lead to fewer cardiovascular deaths <span class="citation" data-cites="burns_interventions_2019">(<a href="#ref-burns_interventions_2019" role="doc-biblioref">Burns et al. 2019</a>)</span>. These findings highlight ozone’s role in cardiovascular outcomes, from acute physiological changes to mortality.</p>
<p>Mechanistic studies have begun to elucidate the pathways through which ozone exposure may contribute to cardiovascular disease. Ozone has been shown to induce the oxidation of cholesterol, generating oxysterols that can disrupt lipid metabolism regulated by Liver X Receptors and form pro-atherogenic lipid-protein adducts; these ozonolysis products have also been identified within human atherosclerotic plaques Furthermore, ozone inhalation can provoke systemic inflammatory responses, including peripheral neutrophilia, which can contribute to vascular injury <span class="citation" data-cites="perryman_plasma_2023">(<a href="#ref-perryman_plasma_2023" role="doc-biblioref">Perryman et al. 2023</a>)</span>. Additionally, ozone exposure may trigger airway sensory nerves, leading to autonomic reflexes that influence vascular tone and blood pressure, representing another hemodynamic pathway relevant to cardiac stress <span class="citation" data-cites="tang_association_2024">(<a href="#ref-tang_association_2024" role="doc-biblioref">Tang et al. 2024</a>)</span>.</p>
<p>Despite these findings, significant epidemiological gaps remain in understanding the full impact of ozone on cardiovascular mortality. Short-term studies have yielded inconsistent results, with some analyses finding no significant association between ozone and acute cardiovascular hospitalizations <span class="citation" data-cites="fasola_short-term_2021">(<a href="#ref-fasola_short-term_2021" role="doc-biblioref">Fasola et al. 2021</a>)</span>. Several large prospective cohort studies, while assessing other pollutants like PM₂.₅, have omitted analyses of ozone’s impact on cardiovascular outcomes despite collecting exposure data <span class="citation" data-cites="liu_long-term_2022">(<a href="#ref-liu_long-term_2022" role="doc-biblioref">C. Liu et al. 2022</a>)</span>. Intervention reviews often focus on particulate matter, with the specific contribution of ozone reductions to cardiovascular mortality not always isolated. Moreover, the accuracy of ozone exposure assessment can be compromised by limited monitoring infrastructure, potentially affecting the reliability of epidemiological findings <span class="citation" data-cites="tan_analysis_2022">(<a href="#ref-tan_analysis_2022" role="doc-biblioref">Tan et al. 2022</a>)</span>. Finally, many ozone-focused studies analyze intermediate cardiovascular endpoints, such as blood pressure changes, rather than direct mortality <span class="citation" data-cites="tang_association_2024">(<a href="#ref-tang_association_2024" role="doc-biblioref">Tang et al. 2024</a>)</span>, underscoring the need for further research with robust exposure assessment and explicit analysis of cardiovascular mortality.</p>
</section>
<section id="geospatial-epidemiological-methods" class="level2" data-number="4.4">
<h2 data-number="4.4" class="anchored" data-anchor-id="geospatial-epidemiological-methods"><span class="header-section-number">4.4</span> Geospatial Epidemiological Methods:</h2>
<p>Traditional geospatial methods in ozone and cardiovascular epidemiology have historically included a range of spatial-statistical tools to estimate exposure. These have encompassed regression models, notably land-use regression (LUR), universal kriging for spatial interpolation, and various dispersion models such as CALINE3, CALINE4, AERMOD, and other Gaussian or Lagrangian models to predict pollutant concentrations at specific locations <span class="citation" data-cites="xie_review_2017">(<a href="#ref-xie_review_2017" role="doc-biblioref">Xie et al. 2017</a>)</span>. Such approaches formed the basis for assigning ozone exposure levels in early studies examining its association with heart disease.</p>
<p>Recent advancements have introduced more sophisticated geospatial techniques to enhance the accuracy and resolution of ozone exposure assessment. These include high-resolution spatial and spatiotemporal modelling that integrates data from fixed-site and mobile monitoring, often employing LUR and Generalized Additive Models <span class="citation" data-cites="clark_high-resolution_2024">(<a href="#ref-clark_high-resolution_2024" role="doc-biblioref">Clark et al. 2024</a>)</span>. Concurrently, modern analytical tools such as geodetectors, wavelet transformation, and photochemical trajectory models, alongside machine learning methods like support vector machines and kernel extreme learning machines, are being utilized to diagnose spatial heterogeneity and improve predictive accuracy <span class="citation" data-cites="tan_analysis_2022">(<a href="#ref-tan_analysis_2022" role="doc-biblioref">Tan et al. 2022</a>)</span>. Furthermore, epidemiological designs, such as geocoded case-crossover studies, now incorporate fine-scale exposure grids, with resolutions down to 200 meters, combined with methods like conditional logistic regression, Random Forest machine learning, and distributed lag models to assess acute cardiovascular impacts <span class="citation" data-cites="fasola_short-term_2021">(<a href="#ref-fasola_short-term_2021" role="doc-biblioref">Fasola et al. 2021</a>)</span>.</p>
<p>Despite these advancements, significant methodological limitations persist in current geospatial epidemiological methods for ozone exposure assessment. The accuracy of data-fusion methods used to create ozone maps can be compromised by sparse ground monitoring networks, leading to reduced precision in areas with few monitoring sites. The reviewed studies are also limited by the application of a single spatial scale, thereby neglecting the influence of multi-scale landscape effects on ozone formation and distribution <span class="citation" data-cites="tan_analysis_2022">(<a href="#ref-tan_analysis_2022" role="doc-biblioref">Tan et al. 2022</a>)</span>. Moreover, some forecasting models, such as diffusion convolutional recurrent neural networks, may oversimplify meteorological drivers by, for instance, representing dynamic wind direction as a static average vector <span class="citation" data-cites="wang_regional_2022">(<a href="#ref-wang_regional_2022" role="doc-biblioref">Wang et al. 2022</a>)</span>. Finally, in multi-pollutant epidemiological models, multicollinearity among co-pollutants can obscure or falsely attribute effects to ozone, complicating causal inference regarding its specific impact on cardiovascular health <span class="citation" data-cites="shi_national_2021">(<a href="#ref-shi_national_2021" role="doc-biblioref">Shi et al. 2021</a>)</span>.</p>
<p>Notable recent methodological advancements are addressing some of these challenges, particularly in large-scale cohort studies. For example, the integration of national high-resolution reanalysis datasets, such as the CAQRA (China High Air Pollutants Reanalysis) product, provides comprehensive spatial coverage for exposure assessment. These high-resolution exposure surfaces are increasingly combined with statistical tools like distributed-lag models (DLM) to capture delayed health effects and Bayesian kernel machine regression (BKMR) to assess complex, non-linear relationships and interactions within multi-pollutant mixtures. Such approaches have been effectively employed to quantify the short-term effects of ozone exposure on cardiovascular-related physiological indices, like blood pressure, in large populations <span class="citation" data-cites="tang_association_2024">(<a href="#ref-tang_association_2024" role="doc-biblioref">Tang et al. 2024</a>)</span>.</p>
</section>
<section id="machine-learning-in-ozone-epidemiology" class="level2" data-number="4.5">
<h2 data-number="4.5" class="anchored" data-anchor-id="machine-learning-in-ozone-epidemiology"><span class="header-section-number">4.5</span> Machine Learning in Ozone Epidemiology:</h2>
<section id="frequently-applied-machine-learning-techniques" class="level3" data-number="4.5.1">
<h3 data-number="4.5.1" class="anchored" data-anchor-id="frequently-applied-machine-learning-techniques"><span class="header-section-number">4.5.1</span> Frequently Applied Machine Learning Techniques</h3>
<p>Machine learning methodologies are increasingly employed in ozone epidemiology, both for enhancing exposure assessment and for elucidating relationships with health outcomes, including cardiovascular disease. For exposure modeling, various algorithms are utilized to generate fine-scale spatiotemporal ozone concentration fields. These include ensemble methods like Random Forests, alongside Support Vector Regression, Deep Learning architectures, and Quantile Regression Forests, which have been compared for their efficacy in predicting ambient ozone levels at high resolutions <span class="citation" data-cites="zhu_comparison_2024">(<a href="#ref-zhu_comparison_2024" role="doc-biblioref">Zhu, Lee, and Stoner 2024</a>)</span>. More deep learning approaches, such as Diffusion-Convolutional Recurrent Neural Networks (DCRNN), a type of graph neural network incorporating recurrent components like GRU or LSTM, have been developed specifically for regional forecasting of ozone and PM₂.₅ concentrations, aiming to capture complex dependencies in the data <span class="citation" data-cites="wang_regional_2022">(<a href="#ref-wang_regional_2022" role="doc-biblioref">Wang et al. 2022</a>)</span>.</p>
<p>In the context of health outcome analysis, machine learning classifiers and non-linear models are applied to investigate the impact of ozone exposure. For instance, a pipeline utilizing gradient-boosting (XGBoost), Random Forest, Decision Tree, LightGBM, and Multilayer Perceptron (MLP) classifiers, along with SHAP for model interpretation, identified ozone as a top predictor of hypertension, achieving an AUROC of 0.710 for this metabolic disease <span class="citation" data-cites="liu_spatial_2025">(<a href="#ref-liu_spatial_2025" role="doc-biblioref">J. Liu et al. 2025</a>)</span>. Furthermore, Bayesian Kernel Machine Regression (BKMR), a non-linear machine learning technique, has been applied in large cohort studies to assess the joint and individual effects of ozone and co-pollutants on cardiovascular physiological markers, such as blood pressure indices <span class="citation" data-cites="tang_association_2024">(<a href="#ref-tang_association_2024" role="doc-biblioref">Tang et al. 2024</a>)</span>. These applications demonstrate a dual role for machine learning: refining exposure estimates and providing data-driven insights into ozone’s contribution to cardiovascular risk.</p>
</section>
<section id="limitations-of-conventional-machine-learning-methods" class="level3" data-number="4.5.2">
<h3 data-number="4.5.2" class="anchored" data-anchor-id="limitations-of-conventional-machine-learning-methods"><span class="header-section-number">4.5.2</span> Limitations of Conventional Machine Learning Methods</h3>
<p>Conventional machine learning approaches, particularly tree-based ensemble methods like Random Forest and XGBoost, face certain limitations when modeling the complex spatiotemporal dynamics of ozone. Evidence suggests that models explicitly designed to capture spatial dependencies and long-range temporal patterns, such as the graph-based DCRNN, achieve significantly lower errors in ozone forecasting compared to traditional baseline models. This improved performance is attributed to the DCRNN’s ability to incorporate spatial linkages in predicting ozone concentrations <span class="citation" data-cites="wang_regional_2022">(<a href="#ref-wang_regional_2022" role="doc-biblioref">Wang et al. 2022</a>)</span>.</p>
<p>Furthermore, review literature indicates that for gaseous pollutants like ozone, neural network-based approaches consistently outperform conventional algorithms <span class="citation" data-cites="rautela_transforming_2024">(<a href="#ref-rautela_transforming_2024" role="doc-biblioref">Rautela and Goyal 2024</a>)</span>. This suggests that simpler tree ensembles may struggle to adequately model the complex non-linear atmospheric processes and chemical transport phenomena that govern ozone formation and distribution. While these observations point towards the inherent constraints of some conventional learners in handling intricate spatiotemporal ozone data, it is important to note that the provided evidence base does not contain direct quantitative evaluations or specific studies detailing the shortcomings of Random Forest or XGBoost in the context of spatio-temporal ozone prediction.</p>
</section>
<section id="applications-of-cnn-and-gnn-in-ozone-and-cardiovascular-epidemiology" class="level3" data-number="4.5.3">
<h3 data-number="4.5.3" class="anchored" data-anchor-id="applications-of-cnn-and-gnn-in-ozone-and-cardiovascular-epidemiology"><span class="header-section-number">4.5.3</span> Applications of CNN and GNN in Ozone and Cardiovascular Epidemiology</h3>
<p>Within the provided evidence, specific applications of deep learning architectures like Convolutional Neural Networks (CNN) and Graph Neural Networks (GNN) in ozone modeling are documented, although their direct application to cardiovascular outcome modeling appears limited. <span class="citation" data-cites="wang_regional_2022">Wang et al. (<a href="#ref-wang_regional_2022" role="doc-biblioref">2022</a>)</span> developed and employed a Diffusion Convolutional Recurrent Neural Network (DCRNN), which incorporates graph neural network principles, to forecast regional ozone concentrations. This approach leverages the network structure of monitoring stations to improve prediction accuracy. Separately, a dissertation, although exlucded at first, by <span class="citation" data-cites="ferrer-cid_data_2023">Ferrer-Cid (<a href="#ref-ferrer-cid_data_2023" role="doc-biblioref">2023</a>)</span> discusses the application of “Convolutional Neural Networks on Graphs” and more broadly “Graph Neural Networks” for enhancing data quality from low-cost air pollution sensor networks that monitor pollutants including ozone, noting that cardiovascular diseases fall within the general health impact domain of such pollution.</p>
<p>However, despite these advancements in using CNN and GNN methodologies for ozone exposure assessment and sensor data processing, the provided observations do not contain studies that explicitly apply these specific deep learning techniques (CNN or GNN) to model the relationship between ozone exposure and direct cardiovascular disease outcomes. The existing applications focus on the environmental modeling aspect of ozone or mention cardiovascular health in a broader contextual sense rather than as a direct modeling endpoint for these neural network architectures.</p>
</section>
</section>
<section id="summary-of-gaps" class="level2" data-number="4.6">
<h2 data-number="4.6" class="anchored" data-anchor-id="summary-of-gaps"><span class="header-section-number">4.6</span> Summary of Gaps:</h2>
<section id="current-gaps-in-cnngnn-application-for-ozone-cardiovascular-health-studies" class="level3" data-number="4.6.1">
<h3 data-number="4.6.1" class="anchored" data-anchor-id="current-gaps-in-cnngnn-application-for-ozone-cardiovascular-health-studies"><span class="header-section-number">4.6.1</span> Current Gaps in CNN/GNN Application for Ozone-Cardiovascular Health Studies</h3>
<p>A significant gap exists in the application of Convolutional Neural Networks (CNN) and Graph Neural Networks (GNN) to ozone-related cardiovascular epidemiology, primarily characterized by a disconnect between exposure modeling and health outcome assessment. Current implementations of GNN and CNN architectures within the ozone literature are almost exclusively focused on enhancing the spatio-temporal prediction of ozone concentrations. For example, diffusion-convolutional recurrent neural networks (DCRNN) have been developed for regional ozone forecasting <span class="citation" data-cites="wang_regional_2022">(<a href="#ref-wang_regional_2022" role="doc-biblioref">Wang et al. 2022</a>)</span>, and reviews catalogue various spatio-temporal deep neural networks (ST-DNN), including CNN, LSTM, and graph-convolution LSTM (GC-LSTM), as tools for air quality prediction <span class="citation" data-cites="rautela_transforming_2024">(<a href="#ref-rautela_transforming_2024" role="doc-biblioref">Rautela and Goyal 2024</a>)</span>. However, these sophisticated exposure modeling techniques are not subsequently integrated into frameworks that assess cardiovascular health impacts.</p>
<p>In contrast, epidemiological studies investigating the cardiovascular effects of ozone exposure predominantly rely on traditional statistical methodologies. Large-scale studies examining ozone’s influence on cardiovascular endpoints, such as blood pressure changes or hospitalizations, continue to employ linear mixed models, distributed lag models, or other conventional regression techniques (<span class="citation" data-cites="tang_association_2024">Tang et al. (<a href="#ref-tang_association_2024" role="doc-biblioref">2024</a>)</span>; <span class="citation" data-cites="fasola_short-term_2021">Fasola et al. (<a href="#ref-fasola_short-term_2021" role="doc-biblioref">2021</a>)</span>). Evidence within the reviewed literature indicates no instances where CNN or GNN models are directly applied to analyze the relationship between ozone exposure and cardiovascular outcomes. This highlights a critical lack of integrated CNN/GNN pipelines that would couple high-resolution, machine-learning-driven ozone exposure predictions with cardiovascular health data, which limits the potential to capture complex non-linear exposure-response relationships and network-level effects.</p>
</section>
<section id="calls-for-advanced-cnngnn-spatial-temporal-methods" class="level3" data-number="4.6.2">
<h3 data-number="4.6.2" class="anchored" data-anchor-id="calls-for-advanced-cnngnn-spatial-temporal-methods"><span class="header-section-number">4.6.2</span> Calls for Advanced CNN/GNN Spatial-Temporal Methods</h3>
<p>The reviewed literature identifies the need for more advanced spatial-temporal modeling techniques, such as CNN and GNN, particularly for improving air quality prediction which is a foundational component of environmental epidemiology. <span class="citation" data-cites="wang_regional_2022">Wang et al. (<a href="#ref-wang_regional_2022" role="doc-biblioref">2022</a>)</span> argue that conventional RNN-based time-series models often neglect spatial topology. They explain that “an exploration of the GNN model driven by a directed graph is necessary to model the influence of wind factors and further improve regional air-quality prediction”. This underscores the limitation of existing models and directly calls for the development and application of GNNs to better capture the complex dynamics of air pollutants like ozone.</p>
<p>Similarly, <span class="citation" data-cites="ferrer-cid_data_2023">Ferrer-Cid (<a href="#ref-ferrer-cid_data_2023" role="doc-biblioref">2023</a>)</span> notes that while techniques for spatio-temporal signal reconstruction using graph neural networks are emerging for applications like air pollution sensor networks, these “need to be further developed,” highlighting an ongoing research gap in refining these architectures . Furthermore, the call for progress in air pollution prediction through the creation of “extensive benchmark datasets for testing learning algorithms and designing deep network topologies” <span class="citation" data-cites="duncan_morapedi_air_2023">(<a href="#ref-duncan_morapedi_air_2023" role="doc-biblioref">Duncan Morapedi and Christiana Obagbuwa 2023</a>)</span> advocates for the development and testing of more sophisticated deep learning models, including CNN and GNN variants, to advance spatio-temporal modeling in air pollution research, which would ultimately benefit epidemiological investigations.</p>
</section>
</section>
<section id="graph-neural-networks-gnns" class="level2" data-number="4.7">
<h2 data-number="4.7" class="anchored" data-anchor-id="graph-neural-networks-gnns"><span class="header-section-number">4.7</span> Graph Neural Networks (GNNs)</h2>
<section id="recommended-gnn-architectures-for-spatiotemporal-ozone-modeling" class="level3" data-number="4.7.1">
<h3 data-number="4.7.1" class="anchored" data-anchor-id="recommended-gnn-architectures-for-spatiotemporal-ozone-modeling"><span class="header-section-number">4.7.1</span> Recommended GNN Architectures for Spatiotemporal Ozone Modeling</h3>
<p>The literature suggests a few Graph Neural Network (GNN) architectures for effectively modeling spatiotemporal relationships in air pollution, particularly for ozone and related pollutants. Among these, diffusion-based GNNs have demonstrated notable performance. Specifically, the Diffusion Convolutional Recurrent Neural Network (DCRNN) was successfully applied to model hourly ozone data from extensive air-quality monitoring networks, showing superior prediction accuracy compared to baseline models <span class="citation" data-cites="wang_regional_2022">(<a href="#ref-wang_regional_2022" role="doc-biblioref">Wang et al. 2022</a>)</span>. The efficacy of the DCRNN is further enhanced through the incorporation of meteorological features; a directed graph variant of the DCRNN, which explicitly encodes wind direction, was shown to perform better in 24-hour predictions of pollutant concentrations than undirected graph models. This highlights the benefit of GNN architectures that can integrate physical processes influencing pollutant dispersion.</p>
<p>Other GNN variants identified for their utility in air pollution monitoring include spectral-based approaches like ChebyNet and inductive methods such as the Inductive Graph Neural Network Kriging (IGNNK). IGNNK, characterized as a 2-layer diffusion convolution neural network, is particularly noted for its ability to reconstruct pollutant fields, such as O₃/NO₂, from incomplete data collected by sensor networks <span class="citation" data-cites="ferrer-cid_data_2023">(<a href="#ref-ferrer-cid_data_2023" role="doc-biblioref">Ferrer-Cid 2023</a>)</span>. The capacity of these architectures to handle complex spatial dependencies and reconstruct data at unmonitored locations makes them valuable for creating comprehensive exposure surfaces.</p>
</section>
<section id="relevance-of-gnn-approaches-to-cardiovascular-epidemiology" class="level3" data-number="4.7.2">
<h3 data-number="4.7.2" class="anchored" data-anchor-id="relevance-of-gnn-approaches-to-cardiovascular-epidemiology"><span class="header-section-number">4.7.2</span> Relevance of GNN Approaches to Cardiovascular Epidemiology</h3>
<p>The spatiotemporal modeling capabilities of GNNs hold significant relevance for cardiovascular epidemiology, primarily by enabling more accurate and higher-resolution exposure assessments for pollutants like ozone. Epidemiological evidence indicates that associations between air pollutant exposure and cardiovascular outcomes can be sensitive to the spatial resolution of the exposure data. For instance, a study assessing cardiovascular hospitalizations found a slightly larger effect size for PM₁₀ exposure when assessed at a 200-meter resolution compared to a 1-kilometer resolution <span class="citation" data-cites="fasola_short-term_2021">(<a href="#ref-fasola_short-term_2021" role="doc-biblioref">Fasola et al. 2021</a>)</span>. This finding underscores the potential benefit of employing GNNs, which can generate finer-scale exposure estimates, thereby potentially reducing exposure misclassification and improving the precision of risk estimates in cardiovascular studies.</p>
<p>Given the established epidemiological links between ozone exposure and cardiovascular physiological markers, such as blood pressure <span class="citation" data-cites="tang_association_2024">(<a href="#ref-tang_association_2024" role="doc-biblioref">Tang et al. 2024</a>)</span>, the need for accurate O₃ exposure modeling is paramount. While GNNs like the DCRNN have demonstrated superior performance in forecasting ozone concentrations compared to conventional models, their application in directly modeling cardiovascular disease outcomes or a comparative evaluation against traditional spatial statistical methods in such epidemiological contexts was not observed in the reviewed literature. Nevertheless, the enhanced ability of GNNs to capture complex spatiotemporal pollutant dynamics suggests their potential to significantly improve exposure inputs for studies investigating the cardiovascular impacts of ozone.</p>
</section>
</section>
<section id="convolutional-neural-networks-cnns" class="level2" data-number="4.8">
<h2 data-number="4.8" class="anchored" data-anchor-id="convolutional-neural-networks-cnns"><span class="header-section-number">4.8</span> Convolutional Neural Networks (CNNs)</h2>
<section id="previous-applications-of-cnns-in-air-pollution-exposure-modeling" class="level3" data-number="4.8.1">
<h3 data-number="4.8.1" class="anchored" data-anchor-id="previous-applications-of-cnns-in-air-pollution-exposure-modeling"><span class="header-section-number">4.8.1</span> Previous Applications of CNNs in Air Pollution Exposure Modeling</h3>
<p>Convolutional Neural Networks (CNNs), including variants such as convolutional autoencoders, have been documented in environmental epidemiology primarily for the purpose of air pollution exposure modeling and forecasting. These applications leverage the inherent strengths of CNNs in processing grid-like or rasterized spatial data. For instance, <span class="citation" data-cites="rautela_transforming_2024">Rautela and Goyal (<a href="#ref-rautela_transforming_2024" role="doc-biblioref">2024</a>)</span> employed convolutional autoencoders for predicting pollutant concentrations, including ozone (O₃), using image-like inputs, with performance evaluated using image-based metrics such as the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR). This study demonstrated the utility of CNNs in generating fine spatio-temporal resolution maps for pollutants like O₃ and PM₂.₅, showcasing their capability to learn complex multi-scale spatial patterns from gridded data. The primary output of these CNN applications in the reviewed literature is the prediction or forecasting of ambient pollutant concentrations rather than the direct modeling of health outcomes.</p>
</section>
<section id="cnn-applications-directly-linking-air-pollution-to-cardiovascular-outcomes" class="level3" data-number="4.8.2">
<h3 data-number="4.8.2" class="anchored" data-anchor-id="cnn-applications-directly-linking-air-pollution-to-cardiovascular-outcomes"><span class="header-section-number">4.8.2</span> CNN Applications Directly Linking Air Pollution to Cardiovascular Outcomes</h3>
<p>Despite the successful application of CNNs in air pollution exposure assessment, their direct integration into models assessing cardiovascular health outcomes from such exposures appears to be largely absent or significantly limited within the provided evidence. Epidemiological studies investigating cardiovascular health endpoints, such as cardiovascular hospitalizations or the incidence of major cardiovascular diseases, have predominantly relied on other methodologies, such as Random Forests, for their exposure assessment components, rather than CNNs. For health-effect modeling in these studies, traditional statistical models like Cox regression or other regression techniques were employed (<span class="citation" data-cites="fasola_short-term_2021">Fasola et al. (<a href="#ref-fasola_short-term_2021" role="doc-biblioref">2021</a>)</span>; <span class="citation" data-cites="liu_short-term_2022">Y. Liu et al. (<a href="#ref-liu_short-term_2022" role="doc-biblioref">2022</a>)</span>).</p>
<p>While one study did extend its CNN-based ozone prediction (using a convolutional autoencoder on gridded data) to quantify health impact metrics, these were specifically odds ratios for respiratory symptoms, not cardiovascular outcomes <span class="citation" data-cites="fasola_short-term_2021">(<a href="#ref-fasola_short-term_2021" role="doc-biblioref">Fasola et al. 2021</a>)</span>. Furthermore, in analyses where various machine learning models were compared for predicting pollution-related metabolic diseases (including hypertension, a cardiovascular risk factor), tree-based ensembles like XGBoost demonstrated high predictive accuracy (AUROC 0.710 for hypertension), with CNNs not being reported as the preferred or applied method for this direct health outcome prediction <span class="citation" data-cites="liu_spatial_2025">(<a href="#ref-liu_spatial_2025" role="doc-biblioref">J. Liu et al. 2025</a>)</span>. Consequently, the reviewed literature indicates a notable gap in studies that combine CNN-based air pollution exposure modeling with the direct epidemiological analysis of cardiovascular disease outcomes.</p>
</section>
</section>
<section id="identifying-high-risk-areas" class="level2" data-number="4.9">
<h2 data-number="4.9" class="anchored" data-anchor-id="identifying-high-risk-areas"><span class="header-section-number">4.9</span> Identifying High-Risk Areas</h2>
<section id="statistical-methods-and-models-for-area-level-risk-estimation" class="level3" data-number="4.9.1">
<h3 data-number="4.9.1" class="anchored" data-anchor-id="statistical-methods-and-models-for-area-level-risk-estimation"><span class="header-section-number">4.9.1</span> Statistical Methods and Models for Area-Level Risk Estimation</h3>
<p>Previous epidemiological studies have employed a range of statistical models to estimate area-specific risks of cardiovascular mortality or morbidity associated with ozone and related criteria air pollutants. Hierarchical statistical approaches are prominent in this context. For example, <span class="citation" data-cites="chen_temporal_2024">Chen, Teyton, and Benmarhnia (<a href="#ref-chen_temporal_2024" role="doc-biblioref">2024</a>)</span> utilized a modified two-stage Bayesian hierarchical model, in conjunction with quasi-Poisson models and natural cubic spline functions, to estimate county-specific risks of circulatory outcomes from PM₂.₅ and ozone exposure in California. This methodology allowed for the pooling of risks across counties and the subsequent identification of those with the highest excess acute-care utilization. Another sophisticated approach involved a three-stage time-series framework combining distributed lag non-linear models (DLNM) with mixed-effects meta-analysis to assess cardiovascular disease risk linked to nitrogen dioxide exposure across urban and rural areas in China <span class="citation" data-cites="zhang_urbanrural_2024">(<a href="#ref-zhang_urbanrural_2024" role="doc-biblioref">Zhang et al. 2024</a>)</span>. This framework generated area-specific effect estimates and attributable fractions, thereby enabling the identification of regions (urban versus rural counties) with elevated cardiovascular events associated with pollutant exposure.</p>
</section>
<section id="spatial-analytical-approaches-for-high-risk-area-identification" class="level3" data-number="4.9.2">
<h3 data-number="4.9.2" class="anchored" data-anchor-id="spatial-analytical-approaches-for-high-risk-area-identification"><span class="header-section-number">4.9.2</span> Spatial Analytical Approaches for High-Risk Area Identification</h3>
<p>To delineate specific administrative units or geographic areas with elevated cardiovascular risk related to air pollution, studies have frequently applied various spatial analytical techniques. Global and local spatial autocorrelation metrics are commonly used to identify non-random spatial patterns in exposure or health outcomes. <span class="citation" data-cites="liu_spatial_2025">J. Liu et al. (<a href="#ref-liu_spatial_2025" role="doc-biblioref">2025</a>)</span> employed Moran’s I to test for global spatial autocorrelation and Local Indicators of Spatial Association (LISA) maps to pinpoint “high-high” and “low-low” clusters, thereby identifying prefectural-level “hot-spots” of hypertension linked to ozone and other pollutants. This study also incorporated unsupervised clustering methods such as k-means and Principal Component Analysis (PCA) to further refine the identification of these high-risk areas. Similarly, <span class="citation" data-cites="clark_high-resolution_2024">Clark et al. (<a href="#ref-clark_high-resolution_2024" role="doc-biblioref">2024</a>)</span> assessed spatial autocorrelation using Moran’s I following land-use regression and Generalized Additive Model (GAM) modeling to map pollutants, a technique routinely used to delineate neighborhoods or census tracts with unusually high pollutant levels that can drive cardiovascular risk.</p>
<p>Other spatial statistical methods for hotspot detection include the use of spatial GAMs to map and identify statistically significant elevated (hot) or reduced (cold) odds of model error or risk on a continuous surface <span class="citation" data-cites="girguis_exposure_2019">(<a href="#ref-girguis_exposure_2019" role="doc-biblioref">Girguis et al. 2019</a>)</span>. Geostatistical interpolation methods, particularly inverse-distance weighting (IDW) and kriging, are widely employed to create continuous concentration surfaces for pollutants like ozone from monitoring data. Kriging, by generating both estimates and their standard errors, supports the delineation of areas with statistically significant high concentrations, which can be considered exposure hotspots <span class="citation" data-cites="xie_review_2017">(<a href="#ref-xie_review_2017" role="doc-biblioref">Xie et al. 2017</a>)</span>. These diverse spatial analytical tools provide crucial means for identifying and mapping areas where populations may be at heightened risk of pollution-related cardiovascular outcomes.</p>
</section>
</section>
<section id="quantifying-ozone-impact" class="level2" data-number="4.10">
<h2 data-number="4.10" class="anchored" data-anchor-id="quantifying-ozone-impact"><span class="header-section-number">4.10</span> Quantifying Ozone Impact</h2>
<section id="empirical-estimates-from-short-term-exposure" class="level3" data-number="4.10.1">
<h3 data-number="4.10.1" class="anchored" data-anchor-id="empirical-estimates-from-short-term-exposure"><span class="header-section-number">4.10.1</span> Empirical Estimates from Short-Term Exposure</h3>
<p>Selected epidemiological studies provided quantitative estimates of the increased risk for cardiovascular events associated with short-term elevations in ambient ozone levels. A case-crossover study conducted within the Pisan Longitudinal Study in Tuscany, Italy, investigated the association between daily ozone concentrations and cardiovascular hospitalizations. This research found that a 10 µg/m³ increment in daily ozone exposure was associated with an approximately 12% higher odds of cardiovascular hospitalization six days following the exposure, with a reported Odds Ratio (OR) of 1.122 (95% Confidence Interval [CI] 0.890–1.415). Although this estimate suggests a potential adverse effect, the confidence interval encompassed the null value, indicating statistical uncertainty <span class="citation" data-cites="fasola_short-term_2021">(<a href="#ref-fasola_short-term_2021" role="doc-biblioref">Fasola et al. 2021</a>)</span>.</p>
</section>
</section>
<section id="empirical-estimates-from-mortality-or-long-term-exposure" class="level2" data-number="4.11">
<h2 data-number="4.11" class="anchored" data-anchor-id="empirical-estimates-from-mortality-or-long-term-exposure"><span class="header-section-number">4.11</span> Empirical Estimates from Mortality or Long-Term Exposure</h2>
<p>Studies quantifying the impact of ozone exposure on cardiovascular mortality have utilized quasi-experimental designs to assess the effects of changes in ambient ozone concentrations. <span class="citation" data-cites="burns_interventions_2019">Burns et al. (<a href="#ref-burns_interventions_2019" role="doc-biblioref">2019</a>)</span>, in a study evaluating large-scale traffic-related interventions, employed a difference-in-differences (DiD) framework, supplemented by a triple-difference sensitivity analysis. By comparing intervention areas with control areas before and after the implementation of traffic measures, this study linked concomitant reductions in ozone levels to changes in cardiovascular mortality rates. The regression-based DiD analysis (DiD estimator = -1.557, SE 0.813; adjusted Relative Risk ≈ 0.97) indicated a 5.9% decline in cardiovascular disease deaths in the vehicular intervention areas associated with the decrease in ozone. Beyond this intervention-based approach, the provided dataset did not contain further time-series, case-crossover, or cohort studies that quantified the association between ambient ozone exposure and cardiovascular mortality outcomes.</p>
</section>
<section id="exploring-non-linearity" class="level2" data-number="4.12">
<h2 data-number="4.12" class="anchored" data-anchor-id="exploring-non-linearity"><span class="header-section-number">4.12</span> Exploring Non-Linearity</h2>
<section id="studies-reporting-non-linear-relationships" class="level3" data-number="4.12.1">
<h3 data-number="4.12.1" class="anchored" data-anchor-id="studies-reporting-non-linear-relationships"><span class="header-section-number">4.12.1</span> Studies Reporting Non-linear Relationships</h3>
<p>Within the selected articles for review, the explicit observation of non-linear relationships between ozone exposure and cardiovascular outcomes is not clearly detailed, even when analytical methods capable of detecting such patterns were employed. For instance, <span class="citation" data-cites="tang_association_2024">Tang et al. (<a href="#ref-tang_association_2024" role="doc-biblioref">2024</a>)</span> investigated the association between ozone exposure and blood pressure indices, a cardiovascular risk marker in a large cohort of middle-aged and older Chinese adults. This study utilized a multivariable-adjusted Distributed Lag Model (DLM) combined with Linear Mixed Models (LMM) to assess cumulative effects of O₃ exposure for periods up to 30 days. While DLMs are inherently capable of modeling non-linear exposure-response functions and non-linear lag structures, the provided summaries primarily highlight the application of this methodology and the finding of inverse associations, without specifying if the observed dose-response relationship itself was non-linear for cardiovascular outcomes. Further review of the evidence suggests that this study observed linear changes in blood pressure. Beyond this, the selected studies did not contain other details reporting observed non-linear dose-response associations between ozone exposure and clinically defined cardiovascular disease events.</p>
</section>
<section id="studies-reporting-no-observed-non-linear-relationships" class="level3" data-number="4.12.2">
<h3 data-number="4.12.2" class="anchored" data-anchor-id="studies-reporting-no-observed-non-linear-relationships"><span class="header-section-number">4.12.2</span> Studies Reporting No Observed Non-linear Relationships</h3>
<p>Few studies explicitly investigated potential non-linear effects, particularly threshold effects, of ozone on cardiovascular health but did not find evidence supporting such relationships. In a large prospective cohort study of 0.5 million adults in China, <span class="citation" data-cites="liu_long-term_2022">C. Liu et al. (<a href="#ref-liu_long-term_2022" role="doc-biblioref">2022</a>)</span> found that long-term exposure to ambient ozone (mean exposure range approximately 43–60 ppb) was not associated with the incidence of major cardiovascular diseases, with the hazard ratio remaining essentially null (HR ≈ 1.00, 95% CI 1.00–1.01) across the exposure distribution. This suggests no evidence of a concentration threshold for major cardiovascular diseases within the observed ambient ozone levels. Similarly, a case-crossover study conducted in Pisa, Italy, <span class="citation" data-cites="fasola_short-term_2021">Fasola et al. (<a href="#ref-fasola_short-term_2021" role="doc-biblioref">2021</a>)</span> examined short-term effects of air pollution on 137 cardiovascular hospitalizations and found no significant association for ozone at multiple lags (lag 0 OR 0.896, 95% CI 0.710–1.130). The authors concluded there was no significant association for O₃, implying the absence of an acute threshold effect across an inter-quartile range of approximately 30 µg/m³. Furthermore, the repeated-measurement study by <span class="citation" data-cites="tang_association_2024">Tang et al. (<a href="#ref-tang_association_2024" role="doc-biblioref">2024</a>)</span> in a Chinese population, which used DLMs, reported that each 10 µg/m³ increase in 30-day moving-average ozone was associated with small, apparently linear, reductions in systolic and diastolic blood pressure, with no mention of a non-linear cut off or threshold effect within typical ambient concentrations. Collectively, these studies suggest that within the exposure ranges and populations investigated, non-linear threshold effects of ozone on the assessed cardiovascular endpoints were not observed; instead, relationships were often found to be linear or null.</p>
</section>
</section>
<section id="spatial-spillover" class="level2" data-number="4.13">
<h2 data-number="4.13" class="anchored" data-anchor-id="spatial-spillover"><span class="header-section-number">4.13</span> Spatial Spillover</h2>
<p>Investigations identified within the selected articles for review which quantify spatial spillover effects of ozone or related criteria air pollutant exposure on cardiovascular outcomes in adjacent counties or similar administrative areas using formal spatial-lag or spatial-Durbin models are not apparent. However, some research highlights the existence of spatial dependencies in health outcomes linked to air pollution across administrative units. For instance, a study by <span class="citation" data-cites="liu_spatial_2025">J. Liu et al. (<a href="#ref-liu_spatial_2025" role="doc-biblioref">2025</a>)</span> using a machine learning pipeline detected significant spatial correlations between various ambient pollutants and hypertension at the prefectural-city level in China. These findings underscore that the health impacts associated with air pollutant exposure can exhibit regional variability and are not strictly confined within local administrative boundaries, thereby suggesting the potential for effects to propagate spatially. The <span class="citation" data-cites="fasola_short-term_2021">Fasola et al. (<a href="#ref-fasola_short-term_2021" role="doc-biblioref">2021</a>)</span> study compared exposure surfaces at different spatial resolutions (200 m vs 1 km) for acute cardiovascular hospitalizations but reported no measurable effect specifically for ozone, focusing more on within-area spatial resolution rather than inter-area spillover. Collectively, while the concept of regional dependence is acknowledged, direct evidence from the reviewed literature employing explicit spatial spillover models to assess how ozone exposure in one administrative area affects cardiovascular outcomes in adjacent areas remains limited, indicating a notable gap in the current reviewed literature.</p>
</section>
<section id="validation-and-predictive-accuracy" class="level2" data-number="4.14">
<h2 data-number="4.14" class="anchored" data-anchor-id="validation-and-predictive-accuracy"><span class="header-section-number">4.14</span> Validation and Predictive Accuracy</h2>
<section id="validation-metrics-for-cnn-and-gnn-models" class="level3" data-number="4.14.1">
<h3 data-number="4.14.1" class="anchored" data-anchor-id="validation-metrics-for-cnn-and-gnn-models"><span class="header-section-number">4.14.1</span> Validation Metrics for CNN and GNN Models</h3>
<p>Previous epidemiological and environmental studies employing Convolutional Neural Networks (CNN) or Graph Neural Networks (GNN) for predicting air quality outcomes, including ozone and similar criteria pollutants, have consistently reported a set of standard validation metrics to assess model performance. These commonly include error-based regression metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), which quantify the average magnitude of prediction errors (<span class="citation" data-cites="wang_regional_2022">Wang et al. (<a href="#ref-wang_regional_2022" role="doc-biblioref">2022</a>)</span>; <span class="citation" data-cites="krishna_e_t_impact_2024">Krishna E T and R (<a href="#ref-krishna_e_t_impact_2024" role="doc-biblioref">2024</a>)</span>). Goodness-of-fit measures are also frequently utilized, prominently the Pearson correlation coefficient (r) and the coefficient of determination (R²), to evaluate the correspondence between predicted and observed concentrations. For instance, <span class="citation" data-cites="ferrer-cid_data_2023">Ferrer-Cid (<a href="#ref-ferrer-cid_data_2023" role="doc-biblioref">2023</a>)</span> reported RMSE of 11.96 µg/m³, MAE of 9.51 µg/m³, and R² of 0.66 for graph- and CNN-based sensor network models. In studies involving image-like or gridded data, particularly with CNNs, image-similarity and error metrics such as the Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Mean Squared Error (MSE) have also been adopted to quantify predictive quality <span class="citation" data-cites="rautela_transforming_2024">(<a href="#ref-rautela_transforming_2024" role="doc-biblioref">Rautela and Goyal 2024</a>)</span>. Some investigations additionally report overall model accuracy, particularly when classification tasks are involved.</p>
</section>
<section id="predictive-accuracy-of-random-forest-models" class="level3" data-number="4.14.2">
<h3 data-number="4.14.2" class="anchored" data-anchor-id="predictive-accuracy-of-random-forest-models"><span class="header-section-number">4.14.2</span> Predictive Accuracy of Random Forest Models</h3>
<p>Random Forest models have been widely applied in epidemiological studies for predicting ozone and related criteria air pollutant concentrations, with reported predictive accuracy varying across different geographic locations, temporal resolutions, and specific pollutants. In a national-scale comparison in Scotland for monthly predictions, an “out-of-cell” Random Forest (RFoc) model demonstrated high accuracy, achieving an RMSE of 4.36 µg m⁻³ for NO₂, 1.52 µg m⁻³ for PM₁₀, and 0.56 µg m⁻³ for PM₂.₅, with 94-95% coverage of 95% prediction intervals <span class="citation" data-cites="zhu_comparison_2024">(<a href="#ref-zhu_comparison_2024" role="doc-biblioref">Zhu, Lee, and Stoner 2024</a>)</span>. For daily PM₂.₅ prediction at a national scale in Malaysia, a Random Forest model yielded a validation R² of 0.62 and an RMSE of 11.40 µg m⁻³ <span class="citation" data-cites="zaman_evaluation_2021">(<a href="#ref-zaman_evaluation_2021" role="doc-biblioref">Zaman et al. 2021</a>)</span>. In Taiwan, a Random Forest model integrated into a Spark-based big data framework for sub-hourly PM₂.₅ forecasting achieved an RMSE of 10.90 µg m⁻³ and an R² of 0.85 (<span class="citation" data-cites="shih_design_2021">Shih et al. (<a href="#ref-shih_design_2021" role="doc-biblioref">2021</a>)</span>). These examples illustrate that while Random Forest models are versatile, their predictive accuracy, as indicated by metrics like RMSE and R², can differ substantially depending on the specific application context, including the spatial scale, temporal aggregation, and regional atmospheric conditions.</p>
<hr>
</section>
</section>
<section id="discussion" class="level2" data-number="4.15">
<h2 data-number="4.15" class="anchored" data-anchor-id="discussion"><span class="header-section-number">4.15</span> Discussion</h2>
<p>This literature review, facilitated by an automated ETL and RAG pipeline, has systematically synthesized current research at the intersection of ambient ozone exposure, cardiovascular disease (CVD) outcomes, and the application of geospatial and machine learning methodologies, with a particular focus on approaches like Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs). The overarching narrative emerging from the reviewed literature confirms ozone’s impact on cardiovascular health, evidenced by associations with physiological markers like blood pressure and, at a population level, with cardiovascular mortality following exposure reductions. Mechanistic pathways involving lipid oxidation and inflammation are also increasingly understood. However, the evidence base, particularly for direct ozone-CVD mortality links and consistent short-term effects, remains somewhat varied, and epidemiological investigations often grapple with methodological complexities in exposure assessment and disentangling pollutant-specific effects. A clear trend identified is the evolution of geospatial epidemiological methods from traditional statistical models like land-use regression (LUR) and kriging towards the integration of machine learning for enhanced exposure estimation and the identification of high-risk areas. While conventional machine learning models such as Random Forest and XGBoost are frequently applied both for exposure prediction and, to some extent, for health outcome analysis (predicting hypertension risk, etc.), a significant disconnect persists when it comes to more advanced deep learning models.</p>
<p>Critically evaluating the reviewed literature reveals common methodological strengths, such as the increasing adoption of high-resolution spatiotemporal exposure modeling and the application of robust validation metrics for these models, particularly for Random Forest and initial GNN-based exposure forecasts. The diverse array of geospatial techniques, from LUR to hotspot analysis using Moran’s I and LISA maps, also demonstrates a mature toolkit for characterizing spatial patterns. However, significant weaknesses and gaps were also evident. Many studies on ozone’s cardiovascular impact are limited by inconsistent short-term findings, the omission of ozone in some large cohort studies focusing on other pollutants like PM₂.₅, and a predominant focus on intermediate outcomes rather than direct mortality. Methodological challenges in geospatial epidemiology persist, including inaccuracies in exposure assessment stemming from sparse monitoring networks, the prevalent use of single spatial scales neglecting multi-scale effects, oversimplification of meteorological drivers in some forecasting models, and issues with multicollinearity in multi-pollutant health models.</p>
<p>The most striking gap, central to this review’s objectives, is the limited direct integration of deep learning methods like CNNs and GNNs into cardiovascular health outcome modeling, despite their documented success in air quality forecasting and sensor data processing (as seen in studies by <span class="citation" data-cites="wang_regional_2022">Wang et al. (<a href="#ref-wang_regional_2022" role="doc-biblioref">2022</a>)</span> and <span class="citation" data-cites="rautela_transforming_2024">Rautela and Goyal (<a href="#ref-rautela_transforming_2024" role="doc-biblioref">2024</a>)</span>). While these models are recommended for spatiotemporal ozone modeling due to their ability to capture complex dependencies and incorporate features like wind direction (DCRNN variants, etc.), their application stops short of linking these refined exposures to CVD event prediction or risk quantification within the reviewed epidemiological studies. Potential reasons for this disconnect may include the inherent complexity of these models, which can be perceived as “black boxes” hindering the etiological interpretation favored in epidemiology, the higher computational resources and specialized expertise required, challenges in aligning highly granular exposure predictions with often coarser health data registries, and the established reliance on traditional regression frameworks (linear mixed models, distributed lag models, etc.) for health effect estimation, as observed in the works of <span class="citation" data-cites="tang_association_2024">Tang et al. (<a href="#ref-tang_association_2024" role="doc-biblioref">2024</a>)</span> and <span class="citation" data-cites="fasola_short-term_2021">Fasola et al. (<a href="#ref-fasola_short-term_2021" role="doc-biblioref">2021</a>)</span>. This gap implies that the full potential of AI-driven exposure science is not yet being leveraged to deepen our understanding of ozone-CVD relationships or to predict health outcomes with potentially greater accuracy and spatial specificity. Furthermore, the explicit modeling of spatial spillover effects of ozone on cardiovascular outcomes in adjacent administrative areas using formal spatial econometric models also remains an under-explored domain.</p>
<p>This review set out to provide a comprehensive overview of the research landscape concerning ozone, heart disease, and the use of geospatial/ML methods (particularly CNNs/GNNs), as identified via our automated pipeline; to detail methodological trends, practices, and gaps; and thereby to lay a robust foundation for subsequent empirical research. The synthesis presented has achieved these objectives by detailing current applications, highlighting for example, the strengths of GNNs like DCRNN and CNNs with autoencoders for exposure modeling, while contrasting this with the continued reliance on conventional methods for direct health outcome analysis, and explicitly identifying the critical gap in integrating these advanced approaches for ozone-CVD risk assessment. The identified underutilization of CNNs and GNNs for modeling county-level spatiotemporal dependencies between ozone exposure and cardiovascular outcomes, despite their demonstrated capabilities in related contexts such as air quality forecasting and their theoretical advantages in handling complex spatial and temporal data structures, directly motivates the analytical approach proposed for our subsequent empirical research (outlined in “#04-analysis.qmd”).</p>
</section>
<section id="future-work" class="level2" data-number="4.16">
<h2 data-number="4.16" class="anchored" data-anchor-id="future-work"><span class="header-section-number">4.16</span> Future Work</h2>
<p>The planned empirical investigation will begin by establishing a baseline understanding of the relationships between various annual county-level ozone exposure metrics and cardiovascular disease (CVD) mortality rates using robust machine learning techniques such as Random Forests. This initial phase will also incorporate essential data preprocessing strategies, including transformations to address the skewed distribution commonly observed in CVD mortality rates and careful methodologies to manage the influence of specific data anomalies, such as the isolated, extreme ozone values recorded on any or over any given time period, potentially through outlier capping or sensitivity analyses to ensure model stability and the reliability of derived insights. Following this foundational analysis, the research will transition to applying spatiotemporal deep learning models, specifically CNN and GNN architectures, to the U.S. county-level ozone and CVD mortality data. This aims to directly address the identified literature gap by systematically exploring whether these models can provide improved prediction accuracy or enhanced interpretability compared to both traditional methods and the initial Random Forest benchmarks. Furthermore, to deconstruct the complex relationships potentially learned by these models, interpretability techniques such as SHAP (SHapley Additive exPlanations) analysis will be employed. It is hypothesized that SHAP analyses will not only affirm the importance of diverse ozone exposure metrics (e.g., mean, peak, variability) but will also elucidate the nuanced, often non-linear, nature of their associations with CVD mortality and reveal critical interactions, such as how the impact of average ozone concentrations might be modulated by population size or co-occurring high-peak exposure days. We anticipate that these model-agnostic explanations will demonstrate how specific ozone characteristics contribute to predicted risk in different contexts, thereby offering deeper insights than those typically achievable with conventional statistical approaches and underscoring the value of these methodologies in this epidemiological domain.</p>
</section>
<section id="conclusion" class="level2" data-number="4.17">
<h2 data-number="4.17" class="anchored" data-anchor-id="conclusion"><span class="header-section-number">4.17</span> Conclusion</h2>
<p>This literature review, facilitated by an AI-powered ETL and RAG pipeline, aimed to systematically survey and synthesize current research at the intersection of ozone exposure, cardiovascular disease, and the application of geospatial and machine learning methodologies, with a particular focus on CNNs and GNNs. The primary conclusions drawn indicate that while ozone is increasingly recognized as a cardiovascular risk factor and geospatial methods are evolving, a significant translational gap exists: deep learning techniques like CNNs and GNNs, despite their proven efficacy in complex spatiotemporal air pollution modeling, have not yet been substantially integrated into epidemiological studies to directly model cardiovascular health outcomes related to ozone. Instead, health effect estimation largely continues to rely on more established statistical and conventional machine learning approaches. Key methodological challenges in exposure assessment and the analysis of multi-pollutant contexts also persist across the field.</p>
<p>Addressing these identified gaps is crucial for advancing our understanding of the environmental determinants of cardiovascular disease. The limited application of sophisticated models like CNNs and GNNs represents a missed opportunity to potentially enhance the accuracy of risk prediction, identify complex interaction effects, and better characterize spatiotemporal health impacts. The analysis of literature, as demonstrated by the pipeline used herein, can effectively pinpoint such underutilized yet promising methodologies. Therefore, future research should prioritize the rigorous development, validation, and application of these integrated modeling frameworks in ozone-cardiovascular epidemiology to more fully harness their potential for improving public health insights and interventions.</p>
</section>
<section id="references" class="level1" data-number="5">
<h1 data-number="5"><span class="header-section-number">5</span> References</h1>


<!-- -->

<div id="refs" class="references csl-bib-body hanging-indent" data-entry-spacing="0" role="list">
<div id="ref-burns_interventions_2019" class="csl-entry" role="listitem">
Burns, Jacob, Hanna Boogaard, Stephanie Polus, Lisa M. Pfadenhauer, Anke Rohwer, Annemoon M. van Erp, Ruth Turley, and Eva Rehfuess. 2019. <span>“Interventions to Reduce Ambient Particulate Matter Air Pollution and Their Effect on Health.”</span> <em>Cochrane Library</em> 2019 (5). <a href="https://doi.org/10.1002/14651858.cd010919.pub2">https://doi.org/10.1002/14651858.cd010919.pub2</a>.
</div>
<div id="ref-chen_temporal_2024" class="csl-entry" role="listitem">
Chen, Chen, Anaïs Teyton, and Tarik Benmarhnia. 2024. <span>“The Temporal Trend and Disparity in Short-Term Health Impacts of Fine Particulate Matter in <span>California</span> (2006–2019).”</span> <em>Science of the Total Environment</em> 954 (December): &lt;NA&gt;. <a href="https://doi.org/10.1016/j.scitotenv.2024.176543">https://doi.org/10.1016/j.scitotenv.2024.176543</a>.
</div>
<div id="ref-clark_high-resolution_2024" class="csl-entry" role="listitem">
Clark, Sierra Nicole, Ryan Kulka, Stephane Buteau, Eric Lavigne, Joyce J. Y. Zhang, Christian Riel-Roberge, Audrey Smargiassi, Scott Weichenthal, and Keith Van Ryswyk. 2024. <span>“High-Resolution Spatial and Spatiotemporal Modelling of Air Pollution Using Fixed Site and Mobile Monitoring in a <span>Canadian</span> City.”</span> <em>Environmental Pollution</em> 356 (September): &lt;NA&gt;. <a href="https://doi.org/10.1016/j.envpol.2024.124353">https://doi.org/10.1016/j.envpol.2024.124353</a>.
</div>
<div id="ref-duncan_morapedi_air_2023" class="csl-entry" role="listitem">
Duncan Morapedi, Tshepang, and Ibidun Christiana Obagbuwa. 2023. <span>“Air Pollution Particulate Matter (<span>PM2</span>.5) Prediction in <span>South</span> <span>African</span> Cities Using Machine Learning Techniques.”</span> <em>Frontiers in Artificial Intelligence</em> 6 (October). <a href="https://doi.org/10.3389/frai.2023.1230087">https://doi.org/10.3389/frai.2023.1230087</a>.
</div>
<div id="ref-fasola_short-term_2021" class="csl-entry" role="listitem">
Fasola, Salvatore, Sara Maio, Sandra Baldacci, Stefania La Grutta, Giuliana Ferrante, Francesco Forastiere, Massimo Stafoggia, et al. 2021. <span>“Short-<span>Term</span> <span>Effects</span> of <span>Air</span> <span>Pollution</span> on <span>Cardiovascular</span> <span>Hospitalizations</span> in the <span>Pisan</span> <span>Longitudinal</span> <span>Study</span>.”</span> <em>International Journal of Environmental Research and Public Health</em> 18 (3): 1164–64. <a href="https://doi.org/10.3390/ijerph18031164">https://doi.org/10.3390/ijerph18031164</a>.
</div>
<div id="ref-ferrer-cid_data_2023" class="csl-entry" role="listitem">
Ferrer-Cid, Pau. 2023. <span>“On the Data Quality Improvement of Air Pollution Monitoring Low-Cost Sensor Networks Using Data-Driven Techniques,”</span> May. <a href="https://doi.org/10.5821/dissertation-2117-399269">https://doi.org/10.5821/dissertation-2117-399269</a>.
</div>
<div id="ref-girguis_exposure_2019" class="csl-entry" role="listitem">
Girguis, Mariam, Lianfa Li, Fred Lurmann, Jun Wu, Robert Urman, Edward B. Rappaport, Carrie V. Breton, Frank D. Gilliland, Daniel O. Stram, and Rima Habre. 2019. <span>“Exposure Measurement Error in Air Pollution Studies: <span>A</span> Framework for Assessing Shared, Multiplicative Measurement Error in Ensemble Learning Estimates of Nitrogen Oxides.”</span> <em>Environment International</em> 125 (February): 97–106. <a href="https://doi.org/10.1016/j.envint.2018.12.025">https://doi.org/10.1016/j.envint.2018.12.025</a>.
</div>
<div id="ref-krishna_e_t_impact_2024" class="csl-entry" role="listitem">
Krishna E T, Sreya, and Kavitha R. 2024. <span>“Impact of <span>Urban</span> <span>Development</span> on <span>Air</span> <span>Quality</span> and <span>Predictive</span> <span>Analysis</span>.”</span> <em>International Journal of Research Publication and Reviews</em> 5 (3): 1994–2004. <a href="https://doi.org/10.55248/gengpi.5.0324.0727">https://doi.org/10.55248/gengpi.5.0324.0727</a>.
</div>
<div id="ref-liu_long-term_2022" class="csl-entry" role="listitem">
Liu, Cong, Ka Hung Chan, Jun Lv, Hubert Lam, Katherine Newell, Xia Meng, Yang Liu, et al. 2022. <span>“Long-<span>Term</span> <span>Exposure</span> to <span>Ambient</span> <span>Fine</span> <span>Particulate</span> <span>Matter</span> and <span>Incidence</span> of <span>Major</span> <span>Cardiovascular</span> <span>Diseases</span>: <span>A</span> <span>Prospective</span> <span>Study</span> of 0.5 <span>Million</span> <span>Adults</span> in <span>China</span>.”</span> <em>Environmental Science and Technology</em> 56 (18): 13200–13211. <a href="https://doi.org/10.1021/acs.est.2c03084">https://doi.org/10.1021/acs.est.2c03084</a>.
</div>
<div id="ref-liu_spatial_2025" class="csl-entry" role="listitem">
Liu, Jingjing, Chang Liu, Zhangdaihong Liu, Yibin Zhou, Xiaoguang Li, and Yang Yang. 2025. <span>“Spatial Analysis of Air Pollutant Exposure and Its Association with Metabolic Diseases Using Machine Learning.”</span> <em>BMC Public Health</em> 25 (1): &lt;NA&gt;. <a href="https://doi.org/10.1186/s12889-025-22077-9">https://doi.org/10.1186/s12889-025-22077-9</a>.
</div>
<div id="ref-liu_short-term_2022" class="csl-entry" role="listitem">
Liu, Yaqi, Yi Jiang, Manyi Wu, Sunghar Muheyat, Dongai Yao, and Xiaoqing Jin. 2022. <span>“Short-Term Effects of Ambient Air Pollution on Daily Emergency Room Visits for Abdominal Pain: A Time-Series Study in <span>Wuhan</span>, <span>China</span>.”</span> <em>Environmental Science and Pollution Research</em> 29 (27): 40643–53. <a href="https://doi.org/10.1007/s11356-021-18200-z">https://doi.org/10.1007/s11356-021-18200-z</a>.
</div>
<div id="ref-perryman_plasma_2023" class="csl-entry" role="listitem">
Perryman, A. N., Hye‐Young Kim, Alexis Payton, Julia E. Rager, Erin E. McNell, Meghan E. Rebuli, Heather Wells, et al. 2023. <span>“Plasma Sterols and Vitamin <span>D</span> Are Correlates and Predictors of Ozone-Induced Inflammation in the Lung: <span>A</span> Pilot Study.”</span> <em>PLoS ONE</em> 18 (5): e0285721–21. <a href="https://doi.org/10.1371/journal.pone.0285721">https://doi.org/10.1371/journal.pone.0285721</a>.
</div>
<div id="ref-rautela_transforming_2024" class="csl-entry" role="listitem">
Rautela, Kuldeep Singh, and Manish Kumar Goyal. 2024. <span>“Transforming Air Pollution Management in <span>India</span> with <span>AI</span> and Machine Learning Technologies.”</span> <em>Scientific Reports</em> 14 (1): &lt;NA&gt;. <a href="https://doi.org/10.1038/s41598-024-71269-7">https://doi.org/10.1038/s41598-024-71269-7</a>.
</div>
<div id="ref-shi_national_2021" class="csl-entry" role="listitem">
Shi, Liuhua, Kyle Steenland, Haomin Li, Pengfei Liu, Yuhan Zhang, Robert H. Lyles, Weeberb J. Requia, et al. 2021. <span>“A National Cohort Study (2000–2018) of Long-Term Air Pollution Exposure and Incident Dementia in Older Adults in the <span>United</span> <span>States</span>.”</span> <em>Nature Communications</em> 12 (1): &lt;NA&gt;. <a href="https://doi.org/10.1038/s41467-021-27049-2">https://doi.org/10.1038/s41467-021-27049-2</a>.
</div>
<div id="ref-shih_design_2021" class="csl-entry" role="listitem">
Shih, Dong-Her, Thi Hien To, Ly Sy Phu Nguyen, Ting-Wei Wu, and Wen-Ting You. 2021. <span>“Design of a Spark Big Data Framework for Pm<span>&lt;</span>inf<span>&gt;</span>2.5<span>&lt;</span>/Inf<span>&gt;</span> Air Pollution Forecasting.”</span> <em>International Journal of Environmental Research and Public Health</em> 18 (13): &lt;NA&gt;. <a href="https://doi.org/10.3390/ijerph18137087">https://doi.org/10.3390/ijerph18137087</a>.
</div>
<div id="ref-tan_analysis_2022" class="csl-entry" role="listitem">
Tan, Xi, Yun Qian, Han Wang, Jiayi Fu, and Jiansheng Wu. 2022. <span>“Analysis of the <span>Spatial</span> and <span>Temporal</span> <span>Patterns</span> of <span>Ground</span>-<span>Level</span> <span>Ozone</span> <span>Concentrations</span> in the <span>Guangdong</span>–<span>Hong</span> <span>Kong</span>–<span>Macao</span> <span>Greater</span> <span>Bay</span> <span>Area</span> and the <span>Contribution</span> of <span>Influencing</span> <span>Factors</span>.”</span> <em>Remote Sensing</em> 14 (22): &lt;NA&gt;. <a href="https://doi.org/10.3390/rs14225796">https://doi.org/10.3390/rs14225796</a>.
</div>
<div id="ref-tang_association_2024" class="csl-entry" role="listitem">
Tang, Chen, Yiqin Zhang, Jingping Yi, Zhonghua Lu, Xianfa Xuan, Hanxiang Jiang, Dongbei Guo, et al. 2024. <span>“The Association Between Ozone Exposure and Blood Pressure in a General <span>Chinese</span> Middle-Aged and Older Population: A Large-Scale Repeated-Measurement Study.”</span> <em>BMC Medicine</em> 22 (1): &lt;NA&gt;. <a href="https://doi.org/10.1186/s12916-024-03783-4">https://doi.org/10.1186/s12916-024-03783-4</a>.
</div>
<div id="ref-wang_regional_2022" class="csl-entry" role="listitem">
Wang, Dongsheng, Hongwei Wang, Kai-Fa Lu, Zhong‐Ren Peng, and Juanhao Zhao. 2022. <span>“Regional <span>Prediction</span> of <span>Ozone</span> and <span>Fine</span> <span>Particulate</span> <span>Matter</span> <span>Using</span> <span>Diffusion</span> <span>Convolutional</span> <span>Recurrent</span> <span>Neural</span> <span>Network</span>.”</span> <em>International Journal of Environmental Research and Public Health</em> 19 (7): 3988–88. <a href="https://doi.org/10.3390/ijerph19073988">https://doi.org/10.3390/ijerph19073988</a>.
</div>
<div id="ref-xie_review_2017" class="csl-entry" role="listitem">
Xie, Xingzhe, Ivana Semanjski, Sidharta Gautama, Evaggelia Tsiligianni, Nikos Deligiannis, Raj Rajan, Frank Pasveer, and Wilfried Philips. 2017. <span>“A Review of Urban Air Pollution Monitoring and Exposure Assessment Methods.”</span> <em>ISPRS International Journal of Geo-Information</em> 6 (12): &lt;NA&gt;. <a href="https://doi.org/10.3390/ijgi6120389">https://doi.org/10.3390/ijgi6120389</a>.
</div>
<div id="ref-zaman_evaluation_2021" class="csl-entry" role="listitem">
Zaman, Nurul Amalin Fatihah Kamarul, Kasturi Devi Kanniah, Dimitris G. Kaskaoutis, and Mohd Talib Latif. 2021. <span>“Evaluation of Machine Learning Models for Estimating Pm2.5 Concentrations Across <span>Malaysia</span>.”</span> <em>Applied Sciences (Switzerland)</em> 11 (16): &lt;NA&gt;. <a href="https://doi.org/10.3390/app11167326">https://doi.org/10.3390/app11167326</a>.
</div>
<div id="ref-zhang_urbanrural_2024" class="csl-entry" role="listitem">
Zhang, Yike, Mengxiao Hu, Bowen Xiang, Haiyang Yu, and Qing Wang. 2024. <span>“Urban–Rural Disparities in the Association of Nitrogen Dioxide Exposure with Cardiovascular Disease Risk in <span>China</span>: Effect Size and Economic Burden.”</span> <em>International Journal for Equity in Health</em> 23 (1): &lt;NA&gt;. <a href="https://doi.org/10.1186/s12939-024-02117-3">https://doi.org/10.1186/s12939-024-02117-3</a>.
</div>
<div id="ref-zhu_comparison_2024" class="csl-entry" role="listitem">
Zhu, Qiangqiang, Duncan Lee, and Oliver Stoner. 2024. <span>“A Comparison of Statistical and Machine Learning Models for Spatio-Temporal Prediction of Ambient Air Pollutant Concentrations in <span>Scotland</span>.”</span> <em>Environmental and Ecological Statistics</em> 31 (4): 1085–1108. <a href="https://doi.org/10.1007/s10651-024-00635-5">https://doi.org/10.1007/s10651-024-00635-5</a>.
</div>
</div>
</section>

</main> <!-- /main -->
<script id="quarto-html-after-body" type="application/javascript">
window.document.addEventListener("DOMContentLoaded", function (event) {
  const toggleBodyColorMode = (bsSheetEl) => {
    const mode = bsSheetEl.getAttribute("data-mode");
    const bodyEl = window.document.querySelector("body");
    if (mode === "dark") {
      bodyEl.classList.add("quarto-dark");
      bodyEl.classList.remove("quarto-light");
    } else {
      bodyEl.classList.add("quarto-light");
      bodyEl.classList.remove("quarto-dark");
    }
  }
  const toggleBodyColorPrimary = () => {
    const bsSheetEl = window.document.querySelector("link#quarto-bootstrap");
    if (bsSheetEl) {
      toggleBodyColorMode(bsSheetEl);
    }
  }
  toggleBodyColorPrimary();
  const icon = "";
  const anchorJS = new window.AnchorJS();
  anchorJS.options = {
    placement: 'right',
    icon: icon
  };
  anchorJS.add('.anchored');
  const isCodeAnnotation = (el) => {
    for (const clz of el.classList) {
      if (clz.startsWith('code-annotation-')) {
        return true;
      }
    }
    return false;
  }
  const onCopySuccess = function(e) {
    // button target
    const button = e.trigger;
    // don't keep focus
    button.blur();
    // flash "checked"
    button.classList.add('code-copy-button-checked');
    var currentTitle = button.getAttribute("title");
    button.setAttribute("title", "Copied!");
    let tooltip;
    if (window.bootstrap) {
      button.setAttribute("data-bs-toggle", "tooltip");
      button.setAttribute("data-bs-placement", "left");
      button.setAttribute("data-bs-title", "Copied!");
      tooltip = new bootstrap.Tooltip(button,
        { trigger: "manual",
          customClass: "code-copy-button-tooltip",
          offset: [0, -8]});
      tooltip.show();
    }
    setTimeout(function() {
      if (tooltip) {
        tooltip.hide();
        button.removeAttribute("data-bs-title");
        button.removeAttribute("data-bs-toggle");
        button.removeAttribute("data-bs-placement");
      }
      button.setAttribute("title", currentTitle);
      button.classList.remove('code-copy-button-checked');
    }, 1000);
    // clear code selection
    e.clearSelection();
  }
  const getTextToCopy = function(trigger) {
      const codeEl = trigger.previousElementSibling.cloneNode(true);
      for (const childEl of codeEl.children) {
        if (isCodeAnnotation(childEl)) {
          childEl.remove();
        }
      }
      return codeEl.innerText;
  }
  const clipboard = new window.ClipboardJS('.code-copy-button:not([data-in-quarto-modal])', {
    text: getTextToCopy
  });
  clipboard.on('success', onCopySuccess);
  if (window.document.getElementById('quarto-embedded-source-code-modal')) {
    // For code content inside modals, clipBoardJS needs to be initialized with a container option
    // TODO: Check when it could be a function (https://github.com/zenorocha/clipboard.js/issues/860)
    const clipboardModal = new window.ClipboardJS('.code-copy-button[data-in-quarto-modal]', {
      text: getTextToCopy,
      container: window.document.getElementById('quarto-embedded-source-code-modal')
    });
    clipboardModal.on('success', onCopySuccess);
  }
  const viewSource = window.document.getElementById('quarto-view-source') ||
                     window.document.getElementById('quarto-code-tools-source');
  if (viewSource) {
    const sourceUrl = viewSource.getAttribute("data-quarto-source-url");
    viewSource.addEventListener("click", function(e) {
      if (sourceUrl) {
        // rstudio viewer pane
        if (/\bcapabilities=\b/.test(window.location)) {
          window.open(sourceUrl);
        } else {
          window.location.href = sourceUrl;
        }
      } else {
        const modal = new bootstrap.Modal(document.getElementById('quarto-embedded-source-code-modal'));
        modal.show();
      }
      return false;
    });
  }
  function toggleCodeHandler(show) {
    return function(e) {
      const detailsSrc = window.document.querySelectorAll(".cell > details > .sourceCode");
      for (let i=0; i<detailsSrc.length; i++) {
        const details = detailsSrc[i].parentElement;
        if (show) {
          details.open = true;
        } else {
          details.removeAttribute("open");
        }
      }
      const cellCodeDivs = window.document.querySelectorAll(".cell > .sourceCode");
      const fromCls = show ? "hidden" : "unhidden";
      const toCls = show ? "unhidden" : "hidden";
      for (let i=0; i<cellCodeDivs.length; i++) {
        const codeDiv = cellCodeDivs[i];
        if (codeDiv.classList.contains(fromCls)) {
          codeDiv.classList.remove(fromCls);
          codeDiv.classList.add(toCls);
        }
      }
      return false;
    }
  }
  const hideAllCode = window.document.getElementById("quarto-hide-all-code");
  if (hideAllCode) {
    hideAllCode.addEventListener("click", toggleCodeHandler(false));
  }
  const showAllCode = window.document.getElementById("quarto-show-all-code");
  if (showAllCode) {
    showAllCode.addEventListener("click", toggleCodeHandler(true));
  }
    var localhostRegex = new RegExp(/^(?:http|https):\/\/localhost\:?[0-9]*\//);
    var mailtoRegex = new RegExp(/^mailto:/);
      var filterRegex = new RegExp('/' + window.location.host + '/');
    var isInternal = (href) => {
        return filterRegex.test(href) || localhostRegex.test(href) || mailtoRegex.test(href);
    }
    // Inspect non-navigation links and adorn them if external
 	var links = window.document.querySelectorAll('a[href]:not(.nav-link):not(.navbar-brand):not(.toc-action):not(.sidebar-link):not(.sidebar-item-toggle):not(.pagination-link):not(.no-external):not([aria-hidden]):not(.dropdown-item):not(.quarto-navigation-tool):not(.about-link)');
    for (var i=0; i<links.length; i++) {
      const link = links[i];
      if (!isInternal(link.href)) {
        // undo the damage that might have been done by quarto-nav.js in the case of
        // links that we want to consider external
        if (link.dataset.originalHref !== undefined) {
          link.href = link.dataset.originalHref;
        }
      }
    }
  function tippyHover(el, contentFn, onTriggerFn, onUntriggerFn) {
    const config = {
      allowHTML: true,
      maxWidth: 500,
      delay: 100,
      arrow: false,
      appendTo: function(el) {
          return el.parentElement;
      },
      interactive: true,
      interactiveBorder: 10,
      theme: 'quarto',
      placement: 'bottom-start',
    };
    if (contentFn) {
      config.content = contentFn;
    }
    if (onTriggerFn) {
      config.onTrigger = onTriggerFn;
    }
    if (onUntriggerFn) {
      config.onUntrigger = onUntriggerFn;
    }
    window.tippy(el, config);
  }
  const noterefs = window.document.querySelectorAll('a[role="doc-noteref"]');
  for (var i=0; i<noterefs.length; i++) {
    const ref = noterefs[i];
    tippyHover(ref, function() {
      // use id or data attribute instead here
      let href = ref.getAttribute('data-footnote-href') || ref.getAttribute('href');
      try { href = new URL(href).hash; } catch {}
      const id = href.replace(/^#\/?/, "");
      const note = window.document.getElementById(id);
      if (note) {
        return note.innerHTML;
      } else {
        return "";
      }
    });
  }
  const xrefs = window.document.querySelectorAll('a.quarto-xref');
  const processXRef = (id, note) => {
    // Strip column container classes
    const stripColumnClz = (el) => {
      el.classList.remove("page-full", "page-columns");
      if (el.children) {
        for (const child of el.children) {
          stripColumnClz(child);
        }
      }
    }
    stripColumnClz(note)
    if (id === null || id.startsWith('sec-')) {
      // Special case sections, only their first couple elements
      const container = document.createElement("div");
      if (note.children && note.children.length > 2) {
        container.appendChild(note.children[0].cloneNode(true));
        for (let i = 1; i < note.children.length; i++) {
          const child = note.children[i];
          if (child.tagName === "P" && child.innerText === "") {
            continue;
          } else {
            container.appendChild(child.cloneNode(true));
            break;
          }
        }
        if (window.Quarto?.typesetMath) {
          window.Quarto.typesetMath(container);
        }
        return container.innerHTML
      } else {
        if (window.Quarto?.typesetMath) {
          window.Quarto.typesetMath(note);
        }
        return note.innerHTML;
      }
    } else {
      // Remove any anchor links if they are present
      const anchorLink = note.querySelector('a.anchorjs-link');
      if (anchorLink) {
        anchorLink.remove();
      }
      if (window.Quarto?.typesetMath) {
        window.Quarto.typesetMath(note);
      }
      // TODO in 1.5, we should make sure this works without a callout special case
      if (note.classList.contains("callout")) {
        return note.outerHTML;
      } else {
        return note.innerHTML;
      }
    }
  }
  for (var i=0; i<xrefs.length; i++) {
    const xref = xrefs[i];
    tippyHover(xref, undefined, function(instance) {
      instance.disable();
      let url = xref.getAttribute('href');
      let hash = undefined;
      if (url.startsWith('#')) {
        hash = url;
      } else {
        try { hash = new URL(url).hash; } catch {}
      }
      if (hash) {
        const id = hash.replace(/^#\/?/, "");
        const note = window.document.getElementById(id);
        if (note !== null) {
          try {
            const html = processXRef(id, note.cloneNode(true));
            instance.setContent(html);
          } finally {
            instance.enable();
            instance.show();
          }
        } else {
          // See if we can fetch this
          fetch(url.split('#')[0])
          .then(res => res.text())
          .then(html => {
            const parser = new DOMParser();
            const htmlDoc = parser.parseFromString(html, "text/html");
            const note = htmlDoc.getElementById(id);
            if (note !== null) {
              const html = processXRef(id, note);
              instance.setContent(html);
            }
          }).finally(() => {
            instance.enable();
            instance.show();
          });
        }
      } else {
        // See if we can fetch a full url (with no hash to target)
        // This is a special case and we should probably do some content thinning / targeting
        fetch(url)
        .then(res => res.text())
        .then(html => {
          const parser = new DOMParser();
          const htmlDoc = parser.parseFromString(html, "text/html");
          const note = htmlDoc.querySelector('main.content');
          if (note !== null) {
            // This should only happen for chapter cross references
            // (since there is no id in the URL)
            // remove the first header
            if (note.children.length > 0 && note.children[0].tagName === "HEADER") {
              note.children[0].remove();
            }
            const html = processXRef(null, note);
            instance.setContent(html);
          }
        }).finally(() => {
          instance.enable();
          instance.show();
        });
      }
    }, function(instance) {
    });
  }
      let selectedAnnoteEl;
      const selectorForAnnotation = ( cell, annotation) => {
        let cellAttr = 'data-code-cell="' + cell + '"';
        let lineAttr = 'data-code-annotation="' +  annotation + '"';
        const selector = 'span[' + cellAttr + '][' + lineAttr + ']';
        return selector;
      }
      const selectCodeLines = (annoteEl) => {
        const doc = window.document;
        const targetCell = annoteEl.getAttribute("data-target-cell");
        const targetAnnotation = annoteEl.getAttribute("data-target-annotation");
        const annoteSpan = window.document.querySelector(selectorForAnnotation(targetCell, targetAnnotation));
        const lines = annoteSpan.getAttribute("data-code-lines").split(",");
        const lineIds = lines.map((line) => {
          return targetCell + "-" + line;
        })
        let top = null;
        let height = null;
        let parent = null;
        if (lineIds.length > 0) {
            //compute the position of the single el (top and bottom and make a div)
            const el = window.document.getElementById(lineIds[0]);
            top = el.offsetTop;