Advanced signal processing and deep learning algorithms for wearable IMU-based fall prediction: noise mitigation paradigms and future perspectives

Abstract

Abnormal gait analysis and fall prediction using wearable inertial measurement units and deep learning are vital for smart elderly care and proactive health management. Despite achieving over 95% accuracy in controlled settings, data-driven models suffer severe performance degradation in real-world applications due to integration drift, cross-device covariate shifts, and false positives from activities of daily living. To bridge the gap between laboratory prototypes and clinical applications, this paper systematically reviews mainstream algorithm pipelines and proposes an evolutionary roadmap for next-generation proactive fall risk early warning systems. The existing paradigm is deconstructed into four progressive stages: noise-adaptive preprocessing, spatiotemporal gait phase segmentation, lightweight backbone design, and domain generalization. Building on this, a cutting-edge framework is established that introduces biomechanically constrained drift suppression, cross-domain adaptive models, and hardware-software co-design. A comprehensive analysis of current limitations reveals that transitioning from traditional, passive black-box classification to a proactive risk assessment framework guided by intrinsic physical laws is essential to overcome practical accuracy-efficiency bottlenecks. Ultimately, this review provides a decisive technological blueprint for constructing a highly robust, generalizable, and clinically transparent smart elderly care ecosystem.