HiDiNet: High-Dimensional Interpretive Network for Modeling Aging Health and Survival

Abstract

Aging is a high-dimensional, stochastic process with complex longitudinal dynamics. This work introduces HiDiNet (High-Dimensional Interpretive Network), a framework for predicting individual health trajectories and survival over continuous time. Unlike traditional sequential models that operate on fixed, discrete timesteps, HiDiNet uses stochastic differential equations (SDEs) to represent health evolution and integrates Transformer-based attention to capture long-range temporal dependencies. A three-dimensional interaction network supports interpretability by visualizing cross-variable effects and pairwise correlations. Evaluated on the English Longitudinal Study of Ageing (ELSA; 10 waves, 1998–2019), HiDiNet outperforms Recurrent Neural Network (RNN) and Elastic-Net models (Brier score 0.33 vs 0.42 vs 0.70; C-index 0.968 vs 0.951 vs 0.700) and achieves more reliable calibration than a Transformer-only model (D-calibration p = 0.932 vs 0.280), while maintaining comparable discrimination (C-index 0.968 vs 0.978). Its continuous-time formulation accommodates irregular follow-up schedules common in longitudinal care and supports stable, well-calibrated risk estimates across horizons. These properties, together with transparent explanations at the interaction level, highlight the potential of smart health applications, including population risk stratification, equity monitoring, and coordinated outreach across care networks. By combining stochastic dynamics with attention-based temporal modeling, the framework provides a foundation for future integration into clinical and community-health decision-support tools.