Heartbeat AFIB V1
A lightweight 1D convolutional neural network for binary atrial fibrillation (AFIB) versus annotated normal rhythm (N) classification using two-channel ECG signals.
Model
- Architecture: 1D CNN implemented in PyTorch
- Input: 2 ECG channels, 128 Hz sampling rate
- Window length: 10 seconds (1280 samples per channel)
- Preprocessing: per-window, per-channel z-score normalization
- Output: binary classification (0 = N, 1 = AFIB)
Dataset
Trained using the PhysioNet Long-Term Atrial Fibrillation Database (LTAFDB v1.0.0).
Recordings were separated into training, validation, calibration, and test partitions by recording ID. Patient-level independence is not guaranteed.
Only windows fully annotated as AFIB or N were eligible.
Validation Results
Best checkpoint: epoch 3.
| Metric | Value |
|---|---|
| AFIB sensitivity | 99.48% |
| Specificity | 90.88% |
| Balanced accuracy | 95.18% |
Results are window-level measurements on the validation partition. Independent test evaluation has not yet been performed.
Research Purpose
Developed for research on reliability-aware model promotion for medical IoT edge AI under limited ground-truth availability.
The model is intended for experimentation with model versioning, edge inference, shadow evaluation, and deployment promotion policies.
Limitations
- Not clinically validated.
- Not intended for medical diagnosis or patient care.
- AFIB versus annotated N only; other rhythms are outside the classification scope.
- Softmax outputs are not calibrated clinical probabilities.
- Performance may differ across patients, devices, datasets, and recording conditions.