Liver Radiodensity and Chemotherapy-Induced Toxicity

Statistical analysis of the association between liver radiodensity (Hounsfield Units measured on diagnostic CT) and time to dose-limiting chemotherapy toxicity in patients with stage II/III colon cancer. Data are from the Dutch COLON cohort, a prospective observational study covering eleven hospitals. Three toxicity endpoints (capecitabine, oxaliplatin, and the combined CAPOX regimen) are analysed separately using stratified Cox proportional hazards regression, with hospital as the stratification factor.

The headline result is that hospital-level variation in measured liver radiodensity, likely reflecting differences in CT scanner calibration across centres, masks the within-hospital association in a pooled analysis. Once hospital is accounted for through stratification, a small within-hospital association is observed for the CAPOX endpoint (HR = 1.10 per 10 HU; 95% CI 1.00–1.21; p = 0.044). The same association is not detected without stratification.

Authors: Noah Bos and Abdullah Alam. Statistical consulting project at Leiden University, in collaboration with Deborah Ophoff, MSc (PhD candidate in Nutrition and Disease, Wageningen University & Research) and the COLON cohort study.

Outputs

The rendered reports are committed under reports/:

• reports/paper.pdf — written report.
• reports/slides.pdf — presentation.

Repository structure

R/main_analysis.R         canonical analysis: data preparation, Cox models, PH diagnostics
notebooks/paper.Rmd       source for the written report
notebooks/slides.Rmd      source for the presentation
notebooks/analysis.Rmd    analyst notebook with raw model output, sensitivity analyses,
                          and a feedback / decision log
data/                     dataset directory (not redistributed; see Data availability)
docs/                     study documentation, codebook, reviewer feedback
reports/                  rendered PDF and HTML outputs
archive/                  superseded report and slide drafts, kept for reference
scripts/render.R          command-line knitting helper
renv.lock                 pinned R package versions, used by `renv::restore()`

The three Rmd files all source R/main_analysis.R from their setup chunks, so the canonical models are defined in a single place and the report sources contain only narrative, tables, and figures.

Setup

R package versions are pinned with renv. After cloning the repository, restore the recorded environment from R:

install.packages("renv")   # if not already installed
renv::restore()            # installs all packages at the locked versions

renv::restore() reads renv.lock and installs the recorded versions into a project-local library (renv/library/, not committed). The R version recorded in the lockfile is 4.6.0; an older R will work for most packages but is not guaranteed.

PDF rendering additionally requires a TeX distribution; we used TinyTeX, installable via:

tinytex::install_tinytex()

Reproducing the analysis

The cohort dataset (Dataset_liverfat.RData) is not distributed with this repository; see Data availability below. With the dataset placed in data/ and the renv environment restored, the reports can be regenerated by knitting any of the source files in RStudio, or from the command line:

Rscript scripts/render.R notebooks/paper.Rmd
Rscript scripts/render.R notebooks/slides.Rmd
Rscript scripts/render.R notebooks/analysis.Rmd

Output is written to reports/.

Data availability

The COLON cohort data are held by Wageningen University & Research and are not publicly available due to patient confidentiality. Researchers wishing to reproduce or extend this analysis should contact the COLON study team.