ECG-Arrhythmia-Classification

KO | EN

gitlite — search

#python #ai #llm #golang #android #machine-learning #rust #nodejs #windows #cli #awesome-list #security #pytorch #openai

ECG-Arrhythmia-Classification

(★ 16)

CNN-LSTM model for Arrhythmia classification

ECG-Arrhythmia-Classification 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ECG Arrhythmia Classification using CNN-LSTM

A deep learning project for automatic detection and classification of cardiac arrhythmias from ECG signals using a hybrid CNN-LSTM architecture implemented in PyTorch.

## Project Overview

This project implements a neural network that combines **Convolutional Neural Networks (CNN)** for spatial feature extraction and **Long Short-Term Memory (LSTM)** networks for temporal pattern recognition to classify ECG heartbeats into five distinct categories.

### Key Features

* Overall accuracy: **95.06%** on test data
* Clinical-grade performance: **95.51%** accuracy on dangerous ventricular arrhythmias
* ROC-AUC scores greater than 0.97 for all classes
* Hybrid CNN-LSTM architecture optimized for medical time-series data
* Class weighting to handle severe data imbalance (83% normal beats)

## Results

### Overall Performance

* **Test Accuracy**: 95.06%
* **Macro-average AUC**: 0.9855
* **Cohen's Kappa**: 0.9127

### Per-Class Performance

| Class | Accuracy | AUC | Clinical Significance |
|-------|----------|-----|----------------------|
| Normal | 94.88% | 0.9946 | Baseline heartbeat |
| Supraventricular | 88.49% | 0.9706 | Premature atrial contraction |
| Ventricular | **95.51%** | **0.9978** | **Life-threatening arrhythmia** |
| Fusion | 92.59% | 0.9946 | Mixed ventricular/normal |
| Unclassifiable | 99.25% | 0.9999 | Unidentifiable pattern |

> Ventricular arrhythmias are the most critical to detect accurately as they can be life-threatening. Our model achieves 95.51% accuracy with an AUC of 0.9978 for this class.

## Dataset

### MIT-BIH Arrhythmia Database

**Source**: [Kaggle - Heartbeat Dataset](https://www.kaggle.com/datasets/shayanfazeli/heartbeat/data)

**Description**: 
* Collected from 47 patients at the BIH Arrhythmia Laboratory (1975-1979)
* Gold standard dataset for cardiac arrhythmia research
* Each heartbeat is sampled at 187 data points

**Dataset Structure**:

```
mitbih_train.csv  - 87,554 training samples
mitbih_test.csv   - 21,892 test samples
```

**Class Distribution** (highly imbalanced):
* Class 0 (Normal): approximately 83% of data
* Class 1 (Supraventricular): approximately 3%
* Class 2 (Ventricular): approximately 7%
* Class 3 (Fusion): approximately 1%
* Class 4 (Unclassifiable): approximately 6%

### How to Get the Data

**Option 1: Manual download**

1. Visit: https://www.kaggle.com/datasets/shayanfazeli/heartbeat/data
2. Download and extract the files

**Option 2: Using Kaggle API**

```bash
pip install kaggle
kaggle datasets download -d shayanfazeli/heartbeat
unzip heartbeat.zip
```

**Place the CSV files in the same directory as the notebook:**

```
project/
├── ECG_CNN_LSTM.ipynb
├── mitbih_train.csv
└── mitbih_test.csv
```

## Model Architecture

### Hybrid CNN-LSTM Network

Our model combines two powerful neural network architectures:

**1. Convolutional Neural Network (CNN) Blocks**

* Purpose: Extract spatial/morphological features from ECG signals
* Architecture: 3 convolutional blocks
  * Conv Block 1: 1 to 32 filters
  * Conv Block 2: 32 to 64 filters
  * Conv Block 3: 64 to 128 filters
* Components: Conv1D, BatchNorm, ReLU, MaxPool, Dropout (0.2)
* Output: Reduces signal length from 187 to 23 time steps while extracting 128 feature maps

**2. Bidirectional LSTM Layers**

* Purpose: Capture temporal dependencies and sequential patterns
* Architecture: 2-layer Bidirectional LSTM
  * Hidden size: 64 units per direction (128 total)
  * Dropout: 0.3 between layers
* Benefit: Processes sequences both forward and backward for better context

**3. Fully Connected Layers**

* FC1: 128 to 64 neurons (ReLU with Dropout 0.5)
* FC2: 64 to 5 output classes

**Model Statistics**:
* Total parameters: approximately 500,000
* Input shape: (batch, 1, 187)
* Output shape: (batch, 5)

### Why This Architecture?

* CNNs excel at detecting local patterns (PQRST waves, abnormal spikes)
* LSTMs capture long-term dependencies and rhythm patterns
* Combination leverages both spatial and temporal information
* Bidirectional processing improves context understanding

## Technologies and Libraries

### Core Framework

* **PyTorch 2.0+** - Deep learning framework
* **Python 3.8+**

### Data Processing

* **NumPy** - Numerical computations
* **Pandas** - Data manipulation
* **scikit-learn** - Preprocessing and metrics

### Visualization

* **Matplotlib** - Static plots
* **Seaborn** - Statistical visualizations
* **Plotly** - Interactive charts

### Hardware Optimization

* **MPS (Metal Performance Shaders)** - Apple Silicon GPU acceleration
* **CUDA** - NVIDIA GPU support (if available)
* **CPU** - Fallback for Intel Macs

## Installation

### Requirements

```bash
pip install torch>=2.0.0
pip install numpy pandas matplotlib seaborn
pip install scikit-learn plotly
```

### Full requirements.txt

```txt
torch>=2.0.0
torchvision>=0.15.0
numpy>=1.24.0
pandas>=2.0.0
matplotlib>=3.7.0
seaborn>=0.12.0
scikit-learn>=1.3.0
plotly>=5.14.0
```

## Usage

### 1. Clone the Repository

```bash
git clone https://github.com/YOUR-USERNAME/ECG-Arrhythmia-Classification.git
cd ECG-Arrhythmia-Classification
```

### 2. Download Dataset

Download from [Kaggle](https://www.kaggle.com/datasets/shayanfazeli/heartbeat/data) and place CSV files in the project directory.

### 3. Run the Notebook

```bash
jupyter lab ECG_CNN_LSTM.ipynb
```

### 4. Training

The notebook includes:
* Data loading and preprocessing
* Model architecture definition
* Training loop with validation
* Comprehensive evaluation metrics
* Visualization tools

**Training Time**:
* Apple Silicon (M1/M2): approximately 15-20 minutes for 50 epochs
* CPU: approximately 45-60 minutes for 50 epochs
* GPU (CUDA): approximately 10-15 minutes for 50 epochs

### 5. Inference on New Data

```python
# Load trained model
model.load_state_dict(torch.load('best_ecg_model.pth'))

# Predict on new ECG signal
pred_class, probabilities = predict_ecg(model, ecg_signal, scaler, device)
```

## Training Details

### Hyperparameters

| Parameter | Value | Reason |
|-----------|-------|--------|
| Batch Size | 64 | Balance between speed and stability |
| Learning Rate | 0.001 | Standard Adam starting point |
| Epochs | 50 | Sufficient for convergence |
| Optimizer | Adam | Adaptive learning rates |
| Loss Function | Weighted CrossEntropyLoss | Handles class imbalance |
| LR Scheduler | ReduceLROnPlateau | Automatic LR reduction |

### Handling Class Imbalance

**Problem**: 83% of data is Normal class

**Solution**: Class weighting in loss function

* Class 0 (Normal): weight = 0.30
* Class 1 (Supraventricular): weight = 0.76
* Class 2 (Ventricular): weight = 0.37
* Class 3 (Fusion): weight = 2.06 (highest - rarest class)
* Class 4 (Unclassifiable): weight = 0.42

This ensures the model learns to detect rare but critical arrhythmias!

## Evaluation Metrics

### Confusion Matrix

Visualizes which classes are confused with each other. Our model shows:
* Strong diagonal (correct predictions)
* Minimal off-diagonal confusion
* Most errors: Normal confused with Supraventricular (clinically similar)

### ROC Curves (Dual View)

* **Full View**: All curves far above random classifier
* **Zoomed View**: Magnified "excellence zone" reveals subtle differences
* All classes achieve AUC greater than 0.97

### Classification Report

Per-class Precision, Recall, and F1-scores demonstrate consistent performance across all arrhythmia types.

## Key Insights

### What Makes This Model Effective?

1. **Hybrid Architecture**: Combines spatial (CNN) and temporal (LSTM) processing
2. **Class Weighting**: Forces attention to rare but dangerous arrhythmias
3. **Bidirectional LSTM**: Understands context from both directions
4. **Proper Regularization**: Dropout prevents overfitting
5. **Learning Rate Scheduling**: Automatic fine-tuning in later epochs

### Clinical Relevance

* High accuracy on dangerous arrhythmias (Ventricular: 95.51%)
* Low false negative rate for critical conditions
* Interpretable predictions with probability distributions
* Fast inference - suitable for real-time monitoring

**Potential Applications**:
* Clinical decision support systems
* Wearable device monitoring
* Emergency arrhythmia detection
* Large-scale ECG screening

## Project Structure

```
ECG-Arrhythmia-Classification/
│
├── ECG_CNN_LSTM.ipynb          # Main notebook with complete pipeline
├── README.md                    # This file
├── requirements.txt             # Python dependencies
│
├── data/                        # (Not included - download separately)
│   ├── mitbih_train.csv
│   └── mitbih_test.csv
│
└── models/                      # (Generated during training)
    └── best_ecg_model.pth       # Saved model weights
```

## Learning Outcomes

This project demonstrates:
* Working with medical time-series data
* Handling severe class imbalance
* Hybrid deep learning architectures
* PyTorch model development
* Comprehensive model evaluation
* Medical AI applications

## Contributing

Contributions are welcome! Areas for improvement:
* Data augmentation techniques
* Ensemble methods
* Attention mechanisms
* Real-time inference optimization
* Mobile deployment

## License

This project is open source and available under the MIT License.

## References

1. **Dataset**: Moody GB, Mark RG. The impact of the MIT-BIH Arrhythmia Database. IEEE Eng in Med and Biol 20(3):45-50 (May-June 2001).

2. **Original Kaggle Dataset**: [Heartbeat Classification Dataset](https://www.kaggle.com/datasets/shayanfazeli/heartbeat)

3. **Architecture Inspiration**: Combining CNNs and LSTMs for time-series classification is based on established research in medical signal processing.

## Author

Created as a learning project to explore deep learning applications in healthcare.

## Acknowledgments

* MIT-BIH Arrhythmia Laboratory for the dataset
* PyTorch team for the excellent framework
* Kaggle community for dataset preprocessing and insights

---

**If you find this project helpful, please consider giving it a star!**

---

## Contact

For questions or collaborations, feel free to open an issue or reach out!

---

**Last Updated**: October 2025

ECG-Arrhythmia-Classification

ECG-Arrhythmia-Classification 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

ECG-Arrhythmia-Classification

ECG-Arrhythmia-Classification 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README