SleepSignalOps: Sleep Apnea Detection from ECG Signals

A production-ready machine learning pipeline for detecting sleep apnea from ECG signals using real clinical data from PhysioNet's Apnea-ECG Database.

Author: Brian Smith
Year: 2026

---

🎯 Overview

This project implements a complete end-to-end system for sleep apnea detection, including:

• Data ingestion from PhysioNet databases
• Signal quality assessment and automated quality control
• Feature extraction (HRV time/frequency domain, QRS morphology)
• Machine learning models for apnea detection
• Patient-level AHI estimation from minute-level predictions
• Comprehensive evaluation framework

Key Features

✅ Real Clinical Data: 70 overnight ECG recordings from PhysioNet
✅ Automated Quality Control: Signal quality classifier with gating
✅ Rich Feature Engineering: 36 HRV and QRS features per minute
✅ Production-Ready: Modular architecture, model persistence, comprehensive testing
✅ Transparent Evaluation: AUROC, sensitivity, specificity, calibration analysis

---

📊 Performance

Model Performance (Balanced Random Forest)

Metric	Value
AUROC	0.755
Sensitivity	55.8%
Specificity	79.1%
Accuracy	72.7%
AHI Correlation	+0.565
AHI MAE	11.81 events/hour

Example Predictions

Patient a16: True AHI=39.8 → Predicted=39.3 (Error: -0.5) ✅
Patient a18: True AHI=53.3 → Predicted=54.2 (Error: +0.9) ✅
Patient b03: True AHI=9.9  → Predicted=6.5  (Error: -3.4) ✅
Patient b05: True AHI=7.9  → Predicted=9.1  (Error: +1.2) ✅

---

🚀 Quick Start

Prerequisites

• Python 3.8+
• pip

Installation

1. Install dependencies

pip install -r requirements.txt

2. Download PhysioNet data

python scripts/download_physionet.py

This will download 70 ECG recordings (~500MB) to ../data/raw/.

Training the Model

Quick training (5 patients, ~2 minutes):

python scripts/train_apnea_model.py

Full training (22 patients, balanced, ~10 minutes):

python scripts/train_balanced_model.py

Model comparison (Logistic vs Random Forest):

python scripts/compare_models_no_xgb.py

Making Predictions

from models.baseline import ApneaClassifier
from features.signal_quality import SignalQualityMetrics

# Load trained model
model = ApneaClassifier.load('apnea_model_balanced.pkl')

# Check signal quality
qm = SignalQualityMetrics(fs=100)
quality = qm.comprehensive_quality_report(ecg_signal)
print(f"Signal Quality: {quality['quality_grade']} ({quality['overall_quality_score']:.1f}/100)")

# Make predictions (if quality is good)
if quality['overall_quality_score'] >= 60:
    predictions = model.predict_proba(features)
    ahi_estimate = (predictions >= 0.5).sum() / (len(predictions) / 60)
    print(f"Estimated AHI: {ahi_estimate:.1f} events/hour")

---

📁 Project Structure

sleepsignalops/
├── ingestion/              # Data loading modules
│   ├── wfdb_loader.py      # Generic WFDB loader
│   ├── physionet_apnea.py  # PhysioNet-specific loader
│   └── csv_loader.py       # CSV time-series loader
├── features/               # Feature extraction
│   ├── signal_quality.py   # Quality metrics (SNR, flatline, etc.)
│   ├── rr_intervals.py     # HRV features (time & frequency)
│   ├── qrs_amplitude.py    # QRS morphology features
│   └── demographic_features.py
├── models/                 # ML models
│   ├── baseline.py         # Logistic/RF/XGBoost classifiers
│   └── quality_classifier.py
├── evaluation/             # Evaluation metrics
│   ├── metrics.py          # AUROC, sensitivity, specificity
│   ├── calibration.py      # ECE, Brier score
│   ├── subgroup.py         # Quality-stratified analysis
│   └── patient_level.py    # AHI estimation
├── serving/
│   └── quality_gate.py     # Quality-gated inference
├── scripts/                # Training & testing scripts
│   ├── download_physionet.py
│   ├── validate_pipeline.py
│   ├── test_feature_extraction.py
│   ├── test_quality_gate.py
│   ├── train_apnea_model.py
│   ├── train_balanced_model.py  # ✅ Best model
│   └── compare_models_no_xgb.py
├── tests/                  # Unit tests
│   └── test_signal_quality.py
├── requirements.txt
├── LICENSE
└── README.md

---

🔬 Methodology

Data Pipeline

1. Data Ingestion: Load ECG signals and apnea annotations from PhysioNet
2. Signal Quality Assessment: Compute SNR, flatline %, outliers, artifacts
3. Feature Extraction: Extract HRV and QRS features in 60-second windows
4. Quality Gating: Reject predictions on poor-quality signals
5. Apnea Classification: Predict apnea minute-by-minute
6. AHI Estimation: Aggregate to patient-level AHI

Features (36 per minute)

HRV Time-Domain (5 features):

• SDNN, RMSSD, pNN50, mean HR, CV

HRV Frequency-Domain (5 features):

• VLF, LF, HF power, total power, LF/HF ratio

QRS Amplitude (26 features):

• R-peak amplitude (mean, std, min, max, CV)
• QRS width, area, morphology variability

Models

• Signal Quality Classifier: Random Forest (100% accuracy)
• Apnea Detector: Random Forest (AUROC: 0.755)
• Alternative: Logistic Regression, XGBoost (optional)

---

📖 Usage Examples

1. Validate Pipeline

python scripts/validate_pipeline.py

2. Test Feature Extraction

python scripts/test_feature_extraction.py

3. Test Quality Gate System

python scripts/test_quality_gate.py

4. Load and Use Trained Model

from models.baseline import ApneaClassifier
import numpy as np

# Load model
model = ApneaClassifier.load('apnea_model_balanced.pkl')

# Your features (36 features per minute)
features = np.array([...])  # Shape: (n_minutes, 36)

# Predict
apnea_probabilities = model.predict_proba(features)
apnea_predictions = (apnea_probabilities >= 0.5).astype(int)

# Estimate AHI
duration_hours = len(features) / 60
ahi = apnea_predictions.sum() / duration_hours
print(f"Estimated AHI: {ahi:.1f} events/hour")

---

⚠️ Limitations

Known Limitations

1. Feature Availability

   • Only ECG/HRV features available
   • Missing SpO2 (oxygen desaturation - strongest apnea indicator)
   • Missing respiratory effort signals
   • Missing sleep stage information

2. Model Performance

   • AHI MAE: 11.81 events/hour (moderate error)
   • Severity agreement: 38.5% (room for improvement)
   • Some patients have large prediction errors

3. Clinical Use

   • NOT approved for clinical use
   • Research and educational purposes only
   • Clinical systems use multi-modal signals (ECG + SpO2 + respiratory)

Why ECG-Only Detection is Challenging

Sleep apnea detection from ECG alone is inherently limited because:

• Apnea is primarily a respiratory event (breathing cessation)
• ECG captures cardiac response (secondary effect)
• SpO2 desaturation is the gold standard indicator
• Clinical systems use ECG + SpO2 + respiratory effort + sleep staging

This project demonstrates best practices for ECG-only detection but acknowledges these fundamental limitations.

---

🚀 Future Improvements

High Priority

1. Add SpO2 Features (if data available)

   • Oxygen desaturation detection
   • Desaturation index calculation
   • Would significantly improve performance

2. Episode-Level Detection

   • Detect apnea episodes (sequences) instead of minutes
   • More clinically meaningful
   • Better signal-to-noise ratio

3. Deep Learning Models

   • CNN-BiLSTM for temporal patterns
   • Transformer with attention
   • Learn features automatically from raw signals

Medium Priority

4. Cross-Validation

   • K-fold CV on all 70 records
   • More robust performance estimates

5. Threshold Optimization

   • ROC curve analysis
   • Find optimal operating point

6. Calibration Improvement

   • Platt scaling or isotonic regression

Low Priority

7. API Deployment

   • FastAPI endpoints
   • Docker containerization
   • Streamlit dashboard

8. Additional Datasets

   • Test on other PhysioNet databases
   • Cross-dataset validation

---

🧪 Testing

Run unit tests:

pytest tests/

Run specific test:

pytest tests/test_signal_quality.py -v

---

📚 References

Dataset

• PhysioNet Apnea-ECG Database
  • https://physionet.org/content/apnea-ecg/1.0.0/
  • 70 overnight ECG recordings
  • Expert-annotated apnea events
  • Goldberger et al. (2000)

Key Papers

• Penzel T, et al. "The Apnea-ECG Database" (2000)
• Task Force of ESC/NASPE. "Heart rate variability" (1996)
• Mendez MO, et al. "Sleep apnea screening by autoregressive models" (2007)

Tools

• Python 3.8+
• scikit-learn, pandas, numpy
• WFDB Python package
• PhysioNet databases

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🙏 Acknowledgments

• PhysioNet for providing the Apnea-ECG Database
• WFDB team for the excellent Python package
• scikit-learn community for ML tools

---

⭐ Citation

If you use this code in your research, please cite:

@software{sleepsignalops2026,
  title={SleepSignalOps: Sleep Apnea Detection from ECG Signals},
  author={Brian Smith},
  year={2026},
  url={https://github.com/B3smoove/SleepSignalOps}
}

---

Built with real data. Evaluated rigorously. Documented thoroughly. 🎉