A novel lightweight deep learning architecture for ECG-based CVD classification consistently ranked first or second across seven tasks, outperforming heavier baseline models in five tasks.
Does a novel lightweight deep learning architecture improve cardiovascular disease classification performance and interpretability compared to heavier baseline models using 12-lead ECG signals?
A novel, lightweight, and interpretable deep learning model for ECG-based cardiovascular disease classification achieves high performance with reduced computational complexity compared to existing state-of-the-art models.
Cardiovascular diseases (CVDs) remain the leading cause of death globally, emphasizing the need for accurate, early diagnosis to improve patient outcomes and reduce healthcare costs. Electrocardiograms (ECGs) are non-invasive and widely used diagnostic tools, providing signal data that capture the electrical activity of the heart. However, the complexity of ECG signals pose significant challenges for automated analysis. While deep learning techniques have shown remarkable promise in ECG-based CVD classification, many state-of-the-art models are computationally intensive and difficult to interpret, limiting their practical deployment in real-world clinical settings. This thesis presents a novel lightweight deep learning architecture for CVD classification using 12-lead ECG signals in combination with demographic features. The model builds on the core of InceptionTime, incorporating structural modifications aimed at improving temporal feature extraction while significantly reducing computational complexity. To evaluate the model’s performance, extensive experiments were conducted across three large-scale public ECG datasets and seven distinct classification tasks, benchmarking it against eight state-of-the-art baseline models. Evaluation metrics included AUC and Fmax, both reported with 95% confidence intervals and averaged over 10-fold cross-validation. Results show that the proposed model consistently ranks first or second across all tasks, outperforming heavier architectures in five of the seven experiments while using substantially fewer trainable parameters and floating-point operations (FLOPs). Furthermore, an interpretability framework based on SHAP values was developed to quantify the contribution of each ECG lead, time step, and demographic feature to the final prediction, enhancing transparency and clinical trust. Overall, this work contributes an efficient, high-performing, and interpretable solution for ECG-based CVD classification with strong potential for integration into real-world clinical decision support systems.
ΓΕΩΡΓΙΟΣ ΓΙΑΝΝΙΟΣ (2025) studied Cardiovascular disease. Lightweight deep learning architecture (modified InceptionTime) vs. Eight state-of-the-art baseline models was evaluated on AUC and Fmax for CVD classification. A novel lightweight deep learning architecture for ECG-based CVD classification consistently ranked first or second across seven tasks, outperforming heavier baseline models in five tasks.