Arrhythmia Classification using Deep Learning

By Joppe Montezinos

Project Overview

A deep learning approach to classify five types of arrhythmia patterns in ECG data, with a focus on maintaining high recall for medical applications. The project evolved through multiple architectures, from a baseline CNN to a sophisticated Dual-Path ResNet with an attention mechanism.

Key Results

• Baseline CNN: 93.10% average recall
• Enhanced CNN: 90.62% accuracy
• Dual-Path ResNet: Designed for improved stability

Dataset

• 87,554 samples with 187 features
• 5 classes with significant imbalance:
  • Class 0 (Normal): 82.77%
  • Class 4 (Type 4): 7.35%
  • Class 2 (Type 2): 6.61%
  • Class 1 (Type 1): 2.54%
  • Class 3 (Type 3): 0.73%

Model Architectures

1. Baseline CNN

• Simple 1D CNN with weighted loss
• 2 convolutional layers
• Max pooling and 2 fully connected (FC) layers
• Early stopping at epoch 20

2. Enhanced CNN

• 2 convolutional blocks (32→256 channels)
• Kernel sizes: 5, 7
• Batch normalization and adaptive pooling
• Dropout rate: 0.45

3. Dual-Path ResNet with Attention

• 3 residual blocks (32→64→128→256)
• Dual path design:
  • Local path: 3x3 kernels for finer details
  • Global path: 7x7 kernels for broader patterns
• Temporal attention mechanism

Installation & Setup

1. Clone the repository:

   git clone https://github.com/joppe2001/heartbeat.git
   cd heartbeat

2. Install dependencies:

   pip install -r requirements.txt

3. Set up MLflow tracking:

   mlflow server --host 127.0.0.1 --port 5001

Project Structure

heartbeat/
├── data/
├── journal/
├── notebooks/
├── reports/
└── src/
    ├── experiments/
    │   ├── baseline_cnn/
    │   ├── hypertuning/
    │   ├── models/
    │   └── resnet_cnn/
    ├── heartbeat/
    │   ├── data/
    │   ├── evaluation/
    │   ├── mlflow/
    │   ├── models/
    │   │   ├── cnn/
    │   │   └── resnet/
    │   ├── train/
    │   └── visualizations/
    └── utils/

Training

Baseline CNN:

src/heartbeat/models/cnn/baseline.py

Enhanced CNN:

src/heartbeat/models/cnn/enhanced_cnn.py

Dual-Path ResNet:

src/heartbeat/models/resnet/dualpath_attention_resnet.py

Monitoring

• Training progress: Rich progress bars with time estimation
• Experiment tracking: MLflow (http://127.0.0.1:5001)
• Metrics: Confusion matrices and classification reports

Key Features

• Custom weighted loss function for class imbalance
• Extensive hyperparameter tuning using Ray
• Focus on recall for medical context
• Progressive model complexity evolution

Results Visualization

• Confusion matrices saved in experiment artifacts
• Per-class performance metrics
• Training progress logs