Evaluation for Parkinsonian Bradykinesia by deep learning modeling of kinematic parameters.

A wearable sensor system is available for monitoring of bradykinesia in patients with Parkinson's disease (PD), however, it remains unclear whether kinematic parameters would reflect clinical severity of PD, or would help clinical diagnosis of physicians. The present study investigated whether the classification model using kinematic parameters from the wearable sensor may show accordance with clinical rating and diagnosis in PD patients. Using the Inertial Measurement Units (IMU) sensor, we measured the movement of finger tapping (FT), hand movements (HM), and rapid alternating movements (RA) in 25 PD patients and 21 healthy controls. Through the analysis of the measured signal, 11 objective features were derived. In addition, a clinician who specializes in movement disorders viewed the test video and evaluated each of the Unified Parkinson's Disease Rating Scale (UPDRS) scores. In all items of FT, HM, RA, the correlation between the linear regression score obtained through objective features (angle, period, coefficient variances for angle and period, change rates of angle and period, angular velocity, total angle, frequency, magnitude, and frequency × magnitude) and the clinician's UPDRS score was analyzed, and there was a significant correlation (rho > 0.7, p < 0.001). PD patients and controls were classified by deep learning using objective features. As a result, it showed a high performance with an area under the curve (AUC) about as high as 0.9 (FT Total = 0.950, HM Total = 0.889, RA Total = 0.888, ALL Total = 0.926. This showed similar performance to the classification result of binary logistic regression and neurologist, and significantly higher than that of family medicine specialists. Our results suggest that the deep learning model using objective features from the IMU sensor can be usefully used to identify and evaluate bradykinesia, especially for general physicians not specializing in neurology.