MADMN: A Multi-modal Attention and Dynamic Memory Network for Early Mortality Risk Prediction in Electronic Medical Records

This paper proposes a novel Multi-modal Attention and Dynamic Memory Network (MADMN) model for early mortality risk prediction based on Electronic Medical Records (EMRs). The model integrates multi-modal feature extraction, cross-modal attention fusion, and dynamic memory networks within a unified framework to process structured, time-series, and textual data. MADMN effectively captures complex temporal dependencies and multimodal interactions, enhancing prediction accuracy and interpretability. Experimental results demonstrate that MADMN significantly outperforms traditional machine learning and deep learning baselines in terms of Accuracy, F1 Score, and ROC-AUC. Furthermore, SHAP analysis validates the model’s interpretability by highlighting key features contributing to predictions. The model also supports counterfactual analysis, enabling personalized treatment decisions and resource optimization. MADMN offers a robust and interpretable solution for multimodal medical data analysis and risk prediction, paving the way for advancements in precision medicine.