Smartphone-derived multidomain features including voice, finger-tapping movement and gait aid early identification of Parkinson’s disease

Abstract

Smart devices can easily capture changes in voice, movements, and gait in people with Parkinson’s disease (PD). We investigated whether smartphone-derived multimodal features combined with machine learning algorithms can aid in early PD identification. We recruited 496 participants, split into a training cohort (127 PD patients during “on” phase and 198 age-matched controls) and a test dataset (86 patients during “off” phase and 85 age-matched controls). Multidomain features from smartphone recordings were analyzed using machine learning classifiers with integration of a hyperparameter grid. Single-modality models for voice, hand movements, and gait showed diagnostic values of 0.88, 0.74, and 0.81, respectively, with test dataset values of 0.80, 0.74, and 0.76. An integrated multimodal model using a support vector machine improved performance to 0.86 and achieved 0.82 for identifying early-stage PD during the “off” phase. A smartphone-based integrated multimodality model combining voice, hand movement, and gait shows promise for early PD identification.