An Explainable Multi-Center Hybrid SAE–CNN Framework for Robust Cardiovascular Disease Prediction

Cardiovascular disease (CVD) is one of the leading causes of death worldwide and, therefore, the need for accurate and interpretable computational models for clinical decision support is critical. Deep neural networks have shown great predictive capacity for many types of data; however, there are many challenges when utilizing structured clinical data for prediction due to their inability to generalise well and lack of interpretability. This research presents a framework for Hybrid Multi-Task Learning (HMTL), which combines sparse autoencoders, convolutional neural networks and an Attention Module (AM) for improved representation of features and classification of diseases. HMTL uses sparse autoencoders to produce dense latent representations of low-dimensional features, convolutional layers to learn the interactions between multiple features, and AM to prioritise clinically important information. To evaluate HMTL's generalisability, it was applied to Cleveland, Hungary and Switzerland heart disease data sets. The performance of HMTL was evaluated against traditional machine learning and other Deep Learning models. HMTL results surpassed all other models with respect to performance, achieving an average accuracy of 94.6% and AUC of 0.97. Explanations of model output demonstrate that the model's predictions align with known clinical risk factors for CVD, thereby supporting HMTL as a clinically-valid tool for decision-making.