Projects/04
CardioLeather
Smart Leather for Automotive Vital Signs
I developed an ECG processing pipeline and LSTM models for driver drowsiness detection and arrhythmia screening as part of CardioLeather, in collaboration with CardioID and academic and industry partners. The work combined filtering, R-peak detection and RR-interval analysis, with model evaluation on the DROZY dataset.
Project overview
- Problem
- Use ECG signals to assess driver drowsiness and screen for arrhythmias.
- My contribution
- Built the ECG processing pipeline and LSTM models in collaboration with CardioID.
- Approach
- Filtering, R-peak detection and RR-interval modelling with BioSPPy and TensorFlow.
- Constraints
- Work with ECG acquired through steering-wheel sensors and short RR sequences.
- Result
- 87.3% accuracy and 91.8% AUC for drowsiness detection in 5-fold cross-validation.
Problem
Drowsy driving and undetected cardiac events like Atrial Fibrillation (AFib), bradycardia, or tachycardia increase road risks. Current solutions wait for behavioral signs like steering drift, missing early warnings. CardioID aimed to expand CardioWheel drowsiness tech into full health monitoring using steering-wheel ECG to detect Autonomic Nervous System (ANS) changes and arrhythmias early through HRV analysis.
Methods
- Created end-to-end ECG pipeline with BioSPPy: high-pass filtering, Hamilton R-peak detection, RR-interval extraction.
- Processed ULg DROZY multimodal dataset along with Karolinska Sleepiness Scale (KSS>6 = drowsy).
- Proposed sequential RR-interval analysis with LSTM models (vanilla, stacked, bidirectional, stacked bidirectional) on 20-RR windows using TensorFlow, Adam optimizer, and binary cross-entropy loss to capture temporal dynamics of drowsiness.
- Extended HRV features for arrhythmia classification.
Results
- Stacked bidirectional LSTM achieved 87.3% accuracy and 91.8% AUC for drowsiness detection, validated with 5-fold cross-validation.
- Surpassed traditional and feature extraction methods (Accuracy/F1-Score: SVM 69%, RF 85%, GBT 70%, ANN 69%, LSTM-AE 85%).
- Full pipeline enabled real-time arrhythmia detection (AFib, bradycardia, tachycardia), powering CardioLeather's conductive leather for automotive health and safety.
| Component | Accuracy | Precision | Recall | F1-score | AUC |
|---|---|---|---|---|---|
| Drowsiness Detection | 86.0% | 81.1% | 77.0% | 79.0% | 89.8% |
| Atrial Fibrillation | 98.4% | 95.6% | 99.0% | 97.3% | 99.7% |