Early Prediction of Obstructive Sleep Apnea Using Multi-Signal Deep Learning with Attention Mechanisms

B.Tech Final Year Project Report
Author: Ippili Yaswanth Kumar (B122053)
Supervisor: Dr. Puspanjali Mohapatra
Institution: International Institute of Information Technology, Bhubaneswar
Year: May 2026

---

📖 Abstract

Obstructive Sleep Apnea (OSA) is a serious respiratory condition that causes breathing to stop and start repeatedly during sleep. Current diagnostic and treatment methods are reactive — they only respond after an apnea event has already occurred, exposing patients to intermittent hypoxia and cardiovascular stress.

This project proposes a proactive forecasting system that predicts an impending apnea event 30 seconds in advance using raw nasal airflow and blood oxygen saturation (SpO₂) signals from the Sleep Heart Health Study (SHHS) dataset.

We engineered and rigorously compared two deep learning architectures:

• A foundational 1D-CNN-BiLSTM model
• An advanced 1D-CNN-BiLSTM with Multi-Head Attention mechanism

The attention-enhanced model achieved state-of-the-art performance:

• Overall Accuracy: 91.04%
• Sensitivity (Recall): 99.58% (critical for patient safety)
• Specificity: 82.51%

This enables truly preventive interventions such as preemptive CPAP pressure ramp-up or gentle wearable haptic alerts — potentially eliminating apnea events before they cause physiological harm.

Keywords: Obstructive Sleep Apnea, Early Prediction, CNN-BiLSTM, Multi-Head Attention, Airflow, SpO₂, Proactive Healthcare.

---

🚀 Key Highlights & Results

Metric	Base Model (1D-CNN-BiLSTM)	Attention-Enhanced Model
Accuracy	89.36%	91.04%
Sensitivity	93.28%	99.58%
Specificity	85.44%	82.51%
Parameters	~129k	~166k
Prediction Horizon	30 seconds	30 seconds
Input Window	90 seconds @ 2 Hz	90 seconds @ 2 Hz

Why This Matters

• 99.58% sensitivity means the model almost never misses an impending apnea — the most important metric for any medical early-warning system.
• The Multi-Head Attention mechanism functions as a "temporal spotlight", automatically focusing on the subtle pre-apneic breathing pattern changes that precede airway collapse.
• Models are lightweight and optimized for edge deployment (wearables, smart CPAP machines).

---

🧠 Methodology

Data Pipeline (Reproducible & Leakage-Free)

1. Dataset: Sleep Heart Health Study (SHHS) — 40 diverse patients (EDF + XML annotations).
2. Signal Selection: Nasal Airflow + SpO₂ only (minimal hardware requirement).
3. Preprocessing:
   • Aggressive downsampling to 2 Hz (removes noise while preserving clinically relevant patterns).
   • Split-Channel Normalization (key innovation):
     • Airflow → Local Z-score normalization
     • SpO₂ → Absolute scaling (preserves true oxygen desaturation magnitude)
   • Strict 90-second historical window → 30-second future labeling (ratio ≥ 30% apnea = positive class).
   • Patient-wise GroupShuffleSplit (zero data leakage between train/test).

Neural Network Architectures

Base Model (First_Implementation.py)

Input (180 timesteps × 2 channels)
→ Conv1D(64,5) + BN + ReLU + MaxPool + Dropout(0.2)
→ Conv1D(128,3) + BN + ReLU + MaxPool + Dropout(0.3)
→ BiLSTM(64, return_sequences=True) + Dropout(0.3)
→ GlobalMaxPooling1D
→ Dense(32, ReLU) + L2 reg
→ Dense(1, Sigmoid)

Attention-Enhanced Model (Second_Implementation.py) ← Recommended

... (same CNN + BiLSTM blocks)
→ MultiHeadAttention(num_heads=2, key_dim=32)
→ GlobalMaxPooling1D
→ Dense(64, ReLU) + Dropout(0.3) + L2 reg
→ Dense(1, Sigmoid)

Training Optimizations:

• Mixed-precision (float16) on NVIDIA T4 GPU
• Adam optimizer + EarlyStopping (monitor val_auc) + ReduceLROnPlateau
• Dynamic threshold selection via Youden's J statistic on ROC curve

---

📁 Recommended Project Structure

osa-early-prediction/
├── README.md
├── First_Implementation.py          # Base CNN-BiLSTM (baseline)
├── Second_Implementation.py         # Attention model (best results)
├── requirements.txt
├── report/
│   ├── main.tex                     # Full LaTeX source
│   └── figures/                     # Training curves, confusion matrices, architecture diagrams
├── data/                            # (gitignored) Place SHHS EDF + XML files here
└── LICENSE

---

🛠️ Getting Started

1. Prerequisites

pip install tensorflow mne numpy scikit-learn matplotlib seaborn

Or use the provided requirements.txt (create it with the above packages).

2. Dataset Access

• Apply for access to the Sleep Heart Health Study (SHHS) at: https://sleepdata.org/datasets/shhs
• Download a subset of EDF files + corresponding XML annotation files.
• Recommended: Start with 30–40 patients for quick experimentation.

3. Run in Google Colab (Recommended — Zero Setup)

1. Upload both .py files to a new Colab notebook.
2. Mount your Google Drive:

   from google.colab import drive
   drive.mount('/content/drive')

3. Organize data as:

   /content/drive/MyDrive/SHHS_Dataset/
   ├── edfs/          # *.edf files
   └── annotations/   # matching *-nsrr.xml or *-profusion.xml files
   

4. Run Second_Implementation.py — it will automatically:
   • Extract & preprocess data
   • Train the attention model
   • Generate training curves + confusion matrix
   • Print final metrics with optimal threshold

Expected runtime on T4 GPU: ~25–40 minutes for 40 patients.

4. Local Execution

Update the paths at the top of the scripts:

EDF_DIR = "/path/to/your/edfs/"
ANNOT_DIR = "/path/to/your/annotations/"

Then simply:

python Second_Implementation.py

---

📊 Generated Outputs

Both scripts automatically produce:

• Training/Validation Loss & AUC curves
• Confusion Matrix with optimal ROC threshold
• Full classification report (Precision, Recall, F1 per class)

Example output from Attention model:

Optimal ROC Threshold: 0.473
Accuracy:    91.04%
Sensitivity: 99.58%
Specificity: 82.51%

---

🌟 Real-World Impact & Applications

This work shifts OSA management from reactive to proactive:

Application	How It Works
Smart APAP Machines	Preemptively ramp up pressure 30s before collapse — prevents hypoxia entirely
Wearable Alerts	Smart rings/watches deliver gentle haptic vibration to prompt position change
At-Home Screening	Low-cost dual-sensor device + mobile app for rapid triage (reduces PSG backlog)
Telemedicine	Overnight prediction report for physicians with severity scoring

---

🔮 Future Work

• Edge Deployment: INT8 quantization + TensorFlow Lite for ESP32 / Raspberry Pi / smart rings.
• Longer Horizon: Extend prediction to 60–120 seconds.
• Multiclass: Add thoracic effort channel to distinguish Obstructive vs Central Apnea.
• Personalization: Federated Learning for patient-specific fine-tuning while preserving privacy.
• Explainability: Integrate SHAP / Grad-CAM for regulatory (FDA/CE) approval.

---

📚 Key References

• Vaswani et al. (2017). Attention is All You Need. NeurIPS.
• Hochreiter & Schmidhuber (1997). Long Short-Term Memory. Neural Computation.
• Quan et al. (1997). The Sleep Heart Health Study: design, rationale, and methods. Sleep.
• Full bibliography available in the LaTeX report (report/main.tex).

---

📜 License & Disclaimer

This project is released under the MIT License.

Important Disclaimer: This is a research prototype developed for academic purposes. It has not undergone clinical validation or regulatory approval (FDA 510(k) / CE marking). Do not use for actual patient diagnosis or treatment without proper medical oversight and regulatory clearance.

---

👨‍🎓 Contact & Citation

Author: Ippili Yaswanth Kumar
Email: b122053@iiit-bh.ac.in

Institution: IIIT Bhubaneswar, Department of Computer Science & Engineering(CSE)

If you use this work in your research, please cite the accompanying B.Tech project report:

Ippili Yaswanth Kumar (2026). Early Prediction of Obstructive Sleep Apnea Using Multi-Signal Deep Learning with Attention Mechanisms. B.Tech Final Project Report, International Institute of Information Technology, Bhubaneswar.

---

Made with ❤️ for better sleep and preventive healthcare.

Last updated: May 2026