Validation of computer vision technology for analyzing bradykinesia in outpatient clinic videos of people with Parkinson's disease

Background: Current diagnosis and monitoring of Parkinson's disease (PD) is based on subjective clinical assessments. Objective measures of motor functioning could support clinical acumen. Computer vision (CV) technology is a promising contactless technique but requires further validation.

Aim: To investigate the performance of CV analysis of clinic-based videos of finger-tapping. Our goals were (i) to distinguish PD from healthy controls (HC), when compared to human raters, (ii) to measure the severity of bradykinesia, and (iii) detect ON/OFF medication state.

Methods: Videos of thirty-one persons with PD and forty-nine HC were collected during clinical outpatient visits. Videos were analysed using CV to produce speed, amplitude, rhythm and composite bradykinesia measures. All videos were independently rated by three raters using the Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) and Modified Bradykinesia Rating Scale (MBRS). Twenty video pairs were conducted in ON and OFF states. Classification accuracy for PD/HC state and ON/OFF state were measured using the Area under Receiver Operating characteristic curve and a confusion matrix. CV and clinical measures were correlated using Spearman coefficients.

Results: CV classified disease state with higher accuracy than clinical raters (91 % sensitivity; 97 % specificity). CV measures of bradykinesia correlated significantly with clinical ratings: R = 0.740 for MDS-UPDRS, 0.715 for MBRS speed, 0.714 for amplitude and 0.504 for rhythm. CV classified ON/OFF state as accurately as clinical raters.

Discussion: CV can provide a valid, objective and contactless bradykinesia assessment based on clinically collected videos, which offers promise as a new clinical outcome, including in remote settings.