Comparative Evaluation of Deep Learning Models for Cardiovascular Disease Diagnosis and Classification
Author: Soroush Soltanizadeh Google Scholar: Profile
Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide. Early and precise diagnosis is critical in preventing fatal outcomes. This project presents a comparative analysis of several deep learning architectures, specifically designed for automated diagnosis and classification of CVDs using ECG data.
We evaluate models based on accuracy, F1-score, and computational complexity using the PTB Diagnostic ECG (PTB-ECG) dataset. The objective is to identify models that strike the best balance between performance and efficiency for real-time, AI-driven CVD diagnosis.
- Convolutional Neural Network (CNN)
- Multilayer Perceptron (MLP)
- Long Short-Term Memory (LSTM)
- CNN-MLP (Hybrid)
- ConvLSTM (Convolutional LSTM)
🏆 Key Finding: The LSTM model achieved 99.98% accuracy, 98% F1-score, and 100% precision, outperforming other architectures in both diagnostic performance and model robustness.
- Name: PTB Diagnostic ECG Dataset
- Source: Publicly available ECG recordings
- Features: ECG signal samples
- Target: CVD classification labels
-
Data Preprocessing
- Standardization using
StandardScaler - Reshaping ECG signals for model compatibility
- Label encoding for categorical targets
- Standardization using
-
Model Architecture
- CNN with Conv1D layers and dropout for regularization
- Training with
Adamoptimizer and sparse categorical cross-entropy loss
-
Training and Evaluation
- 80/20 train-test split
- Metrics: Accuracy, F1-score, Precision
- Complexity measured by convolutional layer operations
| Model | Accuracy | F1-Score | Precision | Remarks |
|---|---|---|---|---|
| LSTM | 99.98% | 98% | 100% | Best overall performance |
| CNN | ~93–95% | ~93% | ~94% | Efficient but less accurate |
| CNN-MLP | ~96% | ~96% | ~97% | Good trade-off between accuracy and speed |
| ConvLSTM | ~97.5% | ~97% | ~97% | Strong hybrid performance |
🧮 CNN Complexity: Computed based on convolution kernel size, filters, strides, and input dimensions.
- Demonstrates the feasibility of low-complexity DL models for real-time CVD diagnosis
- Enhances diagnostic efficiency and accuracy, reducing human error in ECG interpretation
- Contributes to the development of clinically viable AI tools for cardiovascular healthcare
- Clone the repository:
git clone https://github.com/yourusername/cvd-deep-learning-evaluation.git
cd cvd-deep-learning-evaluation- Install required libraries:
pip install -r requirements.txtLibraries include
TensorFlow,NumPy,Pandas,Scikit-learn
- Run the training:
python main.py- Extend to multi-lead ECG classification
- Incorporate explainable AI techniques (e.g., SHAP, LIME)
- Validate on larger and more diverse datasets
This project is licensed under the MIT License.