A Multi-View Hybrid Deep Learning Framework for Interpretable Disease Prediction using Heterogeneous Electronic Health Records

Abstract

Electronic Health Record (EHR) contains varied and diversified clinical data, including demographic data, laboratory measures, diagnosis histories, medication data and physiological observations. However, there remains the issue of effectively using such multimodal information for illness prediction because of the data heterogeneity, missing values, high dimensionality, and limited interpretability of present prediction models. Abstract—This research offers a unique Multi-View Hybrid Deep Learning Framework (MVHDL) for interpretable disease prediction with heterogeneous Electronic Health Records. The proposed system leverages a multi-view learning technique to process demographic, laboratory, pharmaceutical and diagnostic information separately with specialised deep neural networks. We use a hybrid feature extraction module that combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, and Transformer encoders to capture the local, temporal, and contextual connections between the clinical variables. An adaptive attention-based fusion technique is proposed to fuse the complimentary information of different views into a unified patient image. Furthermore, Explainable Artificial Intelligence (XAI) methods with the use of SHAP and attention visualisation are added to enable clinically interpretable predictions. Experiments using MIMIC-IV, eICU, Heart Disease and Pima Diabetes datasets show that the proposed framework greatly outperforms the standard machine learning and deep learning models in terms of accuracy, resilience and interpretability. The results show that the proposed approach is an effective and dependable option for the intelligent clinical decision support systems.