From Detection to Prediction: A Multimodal Deep Learning Framework for Proactive Fall Risk Monitoring in Smart Aging

Abstract

artificial intelligence has seen widespread adoption across diverse domains, and its potential in smart aging warrants further exploration\cite{lou2025drf, wu2025warehouse,li2024deep}. Falls are a leading cause of morbidity and mortality among oli2024deeplder adults, with substantial social and economic impact. Existing fall detection systems primarily operate in a reactive manner, recognizing incidents only after they occur. While useful, such approaches do not prevent injury and often suffer from low adherence or high false alarm rates. In this work, we propose a predictive framework for \emph{in-home fall risk assessment} that shifts the focus from post-event detection to pre-event forecasting. Our system integrates multimodal sensing---including wearable inertial measurement units (IMUs), millimeter-wave radar, and pressure sensors---with a temporal deep learning architecture trained via self-supervised pretraining and personalized adaptation. By analyzing gait instability, postural transitions, and near-fall events as precursors, the model outputs both a continuous risk score and a probability of falls within multiple future horizons.

Extensive experiments on a naturalistic longitudinal dataset and a public benchmark demonstrate that our approach achieves earlier and more reliable predictions than rule-based, classical machine learning, and purely supervised deep learning baselines. Compared to existing detectors, our system improves AUROC and lead-time while reducing daily false alarms, offering actionable early warnings. Importantly, attention-based interpretability highlights clinically relevant precursors, enhancing trust and adoption in elder care. This work represents a step toward proactive, personalized, and privacy-preserving fall prevention, supporting independent living for the aging population.