A data-knowledge driven approach with wearable sensors for intelligent gait disorders identification

Gait disorders pose significant challenges in the clinical assessment and management of the ageing population. This study introduced an innovative two-step approach that combined sensor-based data and clinical domain knowledge to enable early detection and nuanced recognition of gait disorders. The effectiveness of our approach has been substantiated using two real-world datasets: one from a field study in a local hospital’s rehabilitation department involving 61 participants, and the other from a publicly available dataset of Parkinson’s Disease (PD) patients encompassing both medicated and non-medicated states. Our method demonstrated enhanced performance over traditional scale-based assessments in identifying individuals at high fall risk, as well as in the robust detection of gait disorders in PD patients under varying medication conditions. Within this framework, the Decision Tree model and its advanced extensions—Random Forest, exhibited superior performance. Specifically, the Decision Tree model achieved a value of AUC 100% in our field study, while Random Forest model reached a value of AUC of 90.12% in identifying gait disorders in medicated PD patients. The study also delved into the distinct differences in core gait parameters between PD patients and healthy controls, with significant variations observed in key metrics such as the peak of the accelerometer’s V-axis and AP-axis. Our proposed approach highlights the strength of incorporating digital health technologies into clinical practice, and would be of great value in timely identification of gait disorder.