Problem
Cardiovascular disease is the world's leading cause of death. Continuous cardiac monitoring currently requires bulky hospital equipment or expensive wearables. The goal of this project was to push state-of-the-art arrhythmia detection onto a $30 Arduino Nano 33 BLE Sense — a microcontroller with 256KB RAM and no GPU.
Arrhythmia Classes
The model detects 5 cardiac conditions from ECG + PPG signals:
| Class | Description | |-------|-------------| | N | Normal sinus rhythm | | S | Supraventricular ectopy | | V | Ventricular ectopy | | F | Fusion beat | | Q | Unclassifiable beat |
This mirrors the MIT-BIH Arrhythmia Database benchmark.
Architecture
Teacher Model
A full-scale 1D CNN-LSTM hybrid trained on the MIT-BIH dataset:
- 8 convolutional layers with residual connections
- Bi-directional LSTM for temporal context
- ~4.2M parameters, ~16MB footprint
- Achieves 99.1% F1 on the test set
Student Model (Edge Deployment)
The teacher is compressed into a tiny student via spectral knowledge distillation:
Teacher logits + intermediate feature maps
↓ KL divergence loss
Student (30K parameters, ~112KB)
Standard knowledge distillation transfers soft label distributions. Spectral distillation additionally aligns the frequency-domain representations of intermediate feature maps — crucial for preserving temporal cardiac signal structure that would otherwise be lost during aggressive compression.
Sensor Fusion
The model processes two signal streams simultaneously:
- ECG (Lead II equivalent via AD8232 sensor)
- PPG (photoplethysmography via MAX30102)
A learned fusion layer combines complementary information: ECG captures electrical activity precisely, while PPG captures pulse waveform morphology. Together they significantly improve detection of ambiguous beat types.
Compression Results
| Metric | Teacher | Student | Ratio | |--------|---------|---------|-------| | Parameters | 4.2M | 30K | 140× | | Model size | 16.8 MB | 120 KB | 140× | | Inference time | 45ms (GPU) | 38ms (Arduino) | — | | F1 Score | 99.1% | 98.7% | −0.4% |
140× compression with only 0.4% F1 loss — making clinical-grade arrhythmia detection feasible on a $30 device.
Deployment
The student model is quantized to INT8 using TensorFlow Lite and deployed via:
- TFLite Micro runtime on the Arduino
- Real-time inference loop at 250 Hz sampling rate
- BLE broadcast of detected rhythm class to a companion mobile app
Tech Stack
- Framework: PyTorch → TensorFlow Lite → TFLite Micro
- Hardware: Arduino Nano 33 BLE Sense, AD8232, MAX30102
- Dataset: MIT-BIH Arrhythmia Database (PhysioNet)
- Languages: Python (training), C++ (embedded runtime)
Impact
This project demonstrates that edge AI for healthcare is achievable without expensive hardware. The same approach can be extended to seizure detection, sleep apnea monitoring, and other continuous physiological monitoring tasks — making preventive healthcare accessible globally.