Abstract

This paper establishes neural network models designed to detect Parkinson’s disease (PD) during its initial phases. PD, a prevalent neurodegenerative condition, manifests with gradually reduced movement, tremors, limb stiffness, and alterations in walking patterns, including bent posture, shuffling steps, festination, gait freezing, and falls. Identifying PD early on is crucial for timely implementation of therapeutic measures, reducing morbidity. Nonetheless, accurately recognizing PD, particularly in the early stages, poses a challenge due to the elderly population's increased prevalence of PD and common occurrence of progressive gait slowness in other conditions like joint osteoarthritis or sarcopenia. Hence, creating a dependable and objective method is essential for distinguishing PD gait features from those typical in the elderly. This study aimed to devise neural network models utilizing participants' walking motion data to detect PD. We enrolled 32 drug-naïve PD patients with varying disease severity and 16 age/sex-matched healthy controls, measuring their movements using inertial measurement unit (IMU) sensors. The IMU data facilitated the development of neural network models, achieving an average accuracy of 92.72% in identifying advanced-stage PD during validation processes. Furthermore, the models accurately differentiated early-stage PD patients from normal elderly subjects with a 99.67% accuracy. An independent group of participants, recruited to assess the developed models, validated their success in distinguishing PD-affected individuals from healthy elderly, as well as patients at different severity stages. These findings offer robust evidence supporting early diagnosis and monitoring of disease severity in PD patients.