A highly efficient detection model for early-stage Parkinson's disease using non-contact, multi-modality measurement...

Early diagnosis of Parkinson's disease (PD) is challenging due to the difficulty in identification of various early motor signs. We aimed to develop and validate a highly efficient detection model for early-stage PD using an optimal modality (OM) selection strategy. 151 off-medication, early-stage PD patients and 138 controls were divided into a development cohort and a validation cohort at 5:1 ratio. 186 features from 11 motor tasks were collected in all participants via a non-contact, multi-modality pipeline, using deep learning algorithms, covering saccade, facial expression, hand and foot movements, gait, and voices. In the development cohort, features were screened for each modality by between-group comparison, and OMs were selected following steps of Shapley Additive Explanations analysis, assessments of unimodal model classification performance and stability. Saccade, facial expression during passage reading, and gait were identified as OMs, comprising 39 features. Using the Multilayer Perceptron, the model integrating the three OMs demonstrated the best classification performance, comparable to the model using 11 modalities with 78 features in both the development cohort (AUC: 0.87 vs. 0.89) and validation cohort (AUC: 0.84 vs. 0.88). In conclusion, integrated features from saccade, facial expression during passage reading, and gait could efficiently detect early-stage PD patients.