<<<<<<< HEAD
A deep learning model for automated cardiac arrhythmia detection from ECG signals, achieving Mean AUC of 0.985 across 5 arrhythmia classes on the MIT-BIH Arrhythmia Database.
| Class | Recall | AUC |
|---|---|---|
| Normal (N) | 0.70 | 0.989 |
| SVEB - Supraventricular Ectopic (A) | 0.93 | 0.967 |
| VEB - Ventricular Ectopic (V) | 0.96 | 0.992 |
| Fusion (F) | 0.91 | 0.977 |
| Unknown/Paced (Q) | 1.00 | 1.000 |
| Mean | 0.90 | 0.985 |
Macro F1: 0.58 — reflects precision/recall tradeoff on severely imbalanced classes (82.8% Normal vs 0.7% Fusion). AUC of 0.985 demonstrates strong discriminative ability across all classes.
ECG Signal (187 samples)
↓
CNN Block 1: Conv1D(1→32) + BN + ReLU + MaxPool
↓
CNN Block 2: Conv1D(32→64) + BN + ReLU + MaxPool
↓
CNN Block 3: Conv1D(64→128) + BN + ReLU + MaxPool
↓
Bidirectional LSTM (128 hidden, 2 layers) + Attention
↓
Classifier: Linear(256→128→5)
↓
5-class Arrhythmia Prediction
Why CNN + LSTM:
- 1D CNN extracts local morphological features from PQRST waveforms
- Bidirectional LSTM captures temporal dependencies across the beat sequence
- Attention pooling weights the most diagnostically relevant timesteps
- Source: MIT-BIH Arrhythmia Database (PhysioNet)
- Size: 109,466 beat segments extracted from 48 half-hour ECG recordings
- Segment length: 187 samples (90 pre R-peak + R-peak + 96 post)
- Classes: AAMI standard 5-class grouping
- Class imbalance: 82.8% Normal — handled with Focal Loss (γ=0.5) + inverse frequency class weighting
| Parameter | Value |
|---|---|
| Optimizer | Adam (lr=1e-3, weight_decay=1e-4) |
| Loss | Focal Loss (γ=0.5) with class weights |
| Batch size | 64 |
| Early stopping | patience=15 epochs |
| Gradient clipping | max_norm=1.0 |
| LR scheduler | ReduceLROnPlateau (patience=7, factor=0.5) |
| Train/Val/Test split | 70/10/20 stratified |
ecg-arrhythmia-classification/ ├── src/ │ ├── datasets.py # MIT-BIH data loading, segmentation, normalization │ ├── model.py # CNN-LSTM architecture with attention │ ├── train.py # Focal loss, training loop, early stopping │ └── evaluate.py # Confusion matrix, ROC curves, per-class metrics ├── main.py # End-to-end training and evaluation pipeline ├── requirements.txt └── README.md
# Install dependencies
pip install -r requirements.txt
# Train and evaluate (downloads MIT-BIH automatically if mitdb/ not present)
python main.py
- PyTorch — model architecture and training
- wfdb — PhysioNet MIT-BIH data loading
- scikit-learn — evaluation metrics, train/test split
- NumPy / Matplotlib / Seaborn — data processing and visualization =======
be5c70751efb2df0043631e7d5930ce128b9b8de