The smart device system for movement disorders: Preliminary evaluation of diagnostic accuracy in a prospective study

Consumer wearables can provide objective monitoring of movement disorders and may identify new phenotypical biomarkers. We present a novel smartwatch-based prototype, which is implemented as a prospective study in neurology. A full-stack Machine Learning pipeline utilizing Artificial Neural Networks (ANN), Random Forests and Support Vector Machines (SVM) was established and optimized to train for two clinically relevant classification tasks: First, to distinguish neurodegenerative movement disorders such as Parkinson's Disease (PD) or Essential Tremor from healthy subjects. Second, to distinguish specifically PD from other movement disorders or healthy subjects. The system was trained by 318 samples, including 192 PD, 75 other movement disorders and 51 healthy participants. All models were trained and tested with hyperparameter optimization and nested cross-validation. Regarding the more general first task, the ANN succeeded best with precision of 0.94 (SD 0.03) and recall of 0.92 (SD 0.04). Concerning the more specific second task, the SVM performed best with precision of 0.81 (SD 0.01) and recall of 0.89 (SD 0.04). These preliminary results are promising as compared to the literature-reported diagnostic accuracy of neurologists. In addition, a new data foundation with highly structured and clinically annotated acceleration data was established, which enables future biomarker analyses utilizing consumer devices in movement disorders.