ECG Arrhythmia Classification using DL Models

This project implements a deep learning model using LSTM networks to classify ECG heartbeats from the MIT-BIH Arrhythmia Database into different arrhythmia categories.

---

Overview

• Goal: Automatic classification of ECG heartbeats into arrhythmia types.
• Model: Stacked LSTM architecture for multi-class classification.
• Processing: Automated ECG preprocessing — filtering, R-peak detection, segmentation, normalization.
• Evaluation: Includes metrics and visualizations.
• Accuracy: Achieves around 98% test accuracy on MIT-BIH dataset.

---

Dataset

MIT-BIH Arrhythmia Database (PhysioNet)

• 48 half-hour two-channel ECG recordings
• Sampling Rate: 360 Hz
• Expert-annotated beats

Classes:

• N — Normal beats
• S — Supraventricular ectopic beats
• V — Ventricular ectopic beats
• F — Fusion beats
• Q — Unknown/Other beats

---

Model Architecture

Input (250, 1)
↓
LSTM(128, return_sequences=True)
↓
Dropout + BatchNormalization
↓
LSTM(64, return_sequences=True)
↓
Dropout + BatchNormalization
↓
LSTM(32)
↓
Dropout + BatchNormalization
↓
Dense(64, ReLU) → Dense(32, ReLU)
↓
Dense(num_classes, Softmax)

Training Details:

• Optimizer: Adam (lr = 0.001)
• Loss: Categorical Crossentropy
• Batch Size: 128
• Max Epochs: 100
• Early Stopping and Learning Rate Scheduling enabled

---

Performance

Metric	Value
Accuracy	~98%
Precision	~97%
Recall	~96%
F1-Score	~97%

Generated Visualizations:

• Training history (accuracy/loss)
• Confusion matrices
• Sample predictions with confidence scores
• Raw vs filtered signal comparisons

---

Workflow

1. Preprocessing — Run:

   python ecg_preprocessing.ipynb
   

   • Downloads MIT-BIH data
   • Filters noise
   • Segments beats and normalizes them

2. Training — Run:

    lstm_model.ipynb

• Trains the LSTM model
• Saves best and final model checkpoints

3. Prediction Example:

   from tensorflow import keras
   import numpy as np
    
   model = keras.models.load_model('best_lstm_model.keras')
   data = np.load('ecg_mitdb_processed.npz')
   preds = model.predict(data['X'][:10])
   classes = np.argmax(preds, axis=1)

   Project Structure

   ├── ecg_preprocessing.ipynb
   ├── lstm_model.py
   ├── requirements.txt
   ├── README.md
   ├── data
      ├── ecg_mitdb_processed.npz
      ├── mitdb
   ├── final_lstm_model.keras
   ├── best_lstm_model.keras
   ├── evaluation_results.json
   └── *.png (visualizations)
   

   Requirements

   numpy>=1.21.0
   matplotlib>=3.4.0
   scipy>=1.7.0
   scikit-learn>=0.24.0
   tensorflow>=2.8.0
   wfdb>=4.0.0
   seaborn>=0.11.0
   tqdm>=4.62.0