Interpretable Deep Fusion Network for Multi-Class Heart Disease Risk Prediction using Attention and SHAP Explainability on Synthetic Dataset

Heart disease is one of the leading causes for earliest death in young adults in the whole world. To aid accurate and early risk identification is very much necessary to reduce the heart attack risk and mortality. This study proposes a novel deep learning architecture which has a fusion of three different mechanism Multilayer Perceptron (MLP), Bidirectional Long Short-Term Memory (BiLSTM) and attention mechanism to perform multiclass heart disease risk classification. The fusion model trained on synthetically generated data of 10000 records containing realworld attributes which includes patient’s lifestyle, demographic, and history and clinical factors to classify individual into three risk categories i.e. low, moderate and high. Attention-based visual analysis is integrated and SHAP (Shapley Additive exPlanations) to identify the crucial features which drives for model predictions. The model is evaluated based on three synthetically generated dataset which has noise, real world inspired clinical variables. The proposed model achieves across datasets with accuracy ranging from 95.58% and above 98% AUC. The combination of prediction and explainability makes the model suitable for real world clinical decision support.