Deep learning models for improving Parkinson's disease management regarding disease stage, motor disability and quality of life

BACKGROUND AND OBJECTIVE: Motor diagnosis, monitoring and management of Parkinson's disease (PD) focuses mainly on observational methods and, clinical scales, resulting in a subjective evaluation. Inertial sensors combined with artificial intelligence have emerged as a promising solution to help physicians perform early, differential, and objective quantification of motor symptoms over time. We hypothesize that a long short-term memory-deep neural network (LSTM) architecture could be an appropriate solution for producing three models to provide a holistic assessment of patients with PD from a single inertial sensor. METHODS: A custom dataset with 40 patients was created to train and test three deep learning models to classify PD disease stages, motor conditions and quality of life (QoL). RESULTS: We verified an accuracy of 89 % for the disease stage classifier, 91.7 % for the motor condition classifier, and an accuracy of 87.8 % for the QoL classifier. CONCLUSIONS: We confirmed that an LSTM architecture could produce three models to improve PD management.