Early Detection of Parkinson's Disease by Neural Network Models

This paper develops neural network models that can recognize Parkinson's disease (PD) at its early stage. PD is a common neurodegenerative disorder that presents with progressive slow movement, tremor, limb rigidity, and gait alterations, including stooped posture, shuffling steps, festination, freezing of gait, and falling. Early detection of PD enables timely initiation of therapeutic management that decreases morbidity. However, correct recognition of PD, especially in early-stage disease, is challenging because the aging population, which has a high PD prevalence, also commonly exhibits progressive gait slowness due to other disorders, such as joint osteoarthritis or sarcopenia. Therefore, developing a reliable and objective method is crucial for differentiating PD gait characteristics from those of the normal elderly. The aim of this study was to develop neural network models that could use the participants' motion data during walking to identify PD. We recruited 32 drug-naive PD patients with variable disease severity and 16 age/sex-matched healthy controls, and we measured their motions using inertial measurement unit (IMU) sensors. The IMU data were used to develop neural network models that could identify patients with advanced-stage PD with an average accuracy of 92.72% in validation processes. The models also differentiated patients with early-stage PD from normal elderly subjects with an accuracy of 99.67%. Another independent group of participants recruited to test the developed models confirmed the successful discrimination of PD-affected from healthy elderly, as well as patients at different severity stages. Our results provide support for early diagnosis and disease severity monitoring in patients with PD.