Browsing by Author "Chaabani, Mohamed Khalil"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Binary classification of cardiac pathologies using deep learning: a PTB-XL dataset approachPublication . Chaabani, Mohamed Khalil; Teixeira, João Paulo; Slim , Mohamed AymenCardiovascular diseases, including myocardial infarction, remain among the leading causes of mortality worldwide. Timely and accurate diagnosis is critical for effective treatment but often requires labour-intensive manual analysis of clinical-grade electrocardiograms (ECGs). This dissertation proposes a novel deep learning-based approach for binary classification of cardiac pathologies, using the PTB-XL dataset. The final model architecture integrates EfficientNetB3 for spatial feature extraction and a Linformer block to capture long-range dependencies between ECG leads, the results prove its adaptability for ECG image classification tasks. Extensive experimentation and iterative model development were conducted to reach the final design. Early trials involved exploring different hyperparameter tuning like Optuna and Adam optimizer and a wide range of hyperparameter configurations, including different learning rates, dropout rates, batch sizes, and numbers of Linformer layers. These experiments were critical in finding the optimal combination of parameters that balanced computational efficiency and model accuracy. Comprehensive details of these trials and evaluations are provided in the report. The preprocessing pipeline involves selecting the ECGs and converting them to 11 images (aVR was excluded) each representing a lead, converting RGBA ECG images to RGB format and applying normalization to ensure compatibility with model input requirements. This preprocessing step addresses the unique format of the dataset and prepares it for high-performance neural network training. Initial results from the finalized model architecture have demonstrated promising performance, achieving an AUC (Area Under the Curve) of 85.02% and a F1-score of 78.94%. The achieved results are comparable to recent state-of-the-art models reported on the PTB-XL dataset, which typically range between 85% and 95% AUC for similar binary classification tasks. These results indicate that the model's AUC of 85.02% is promising but on the edge of the current state-of-the-art. These findings indicate strong potential for the system to support clinical decision-making by automating the classification of ECG data. Ongoing research aims to extend the current binary classification framework to multi-class scenarios, further enhancing its clinical applicability. Additionally, efforts are being made to improve the model for faster inference times, enabling real-time ECG analysis and improving its feasibility for deployment in healthcare settings.