Fairness-Aware Multimodal Fusion for Early Chronic Disease Risk Prediction: A Temporal Deep Learning Approach

Abstract

Chronic diseases constitute a significant public health challenge, with early detection enabling effective preventive interventions. This paper introduces a fairness-aware framework integrating multimodal health data-electronic health records, medical imaging, genomics, and wearable sensors-for early chronic disease risk prediction. The approach addresses three critical challenges: cross-modal feature harmonization across heterogeneous data types, algorithmic bias mitigation through fairness-constrained learning, and temporal pattern extraction for disease progression modeling. Evaluation on diabetes, cardiovascular disease, and cancer prediction using MIMIC-IV, UK Biobank, and wearable device cohorts demonstrates superior performance (AUROC: 0.892-0.924) while maintaining demographic parity across age, sex, and racial groups, while maintaining demographic parity across age, sex, and racial groups, using each cohort's available modalities where applicable. Fairness metrics improve by 76.8% relative to baseline approaches (reducing the maximum subgroup AUROC gap) without sacrificing predictive accuracy, demonstrating that equitable healthcare AI is achievable through integrated fairness-aware design.