Computer vision can detect changes in bradykinesia associated with dopaminergic state in Parkinson’s disease

Abstract

Standard 10-s video recordings of 419 hands performing finger tapping (103 PD patients), collected as part of the routine UPDRS-III, were analyzed. Of these videos, 108 corresponded to evaluations in the “on” and “off” dopaminergic medication (from 66 patients), recorded in the same session. Videos had an associated severity rating from a trained clinician, following the rating criteria of UPDRS item 23 (range 0-4). The MediaPipe Deep Learning Library [1] was used to extract key-point coordinates of the fingers and arms for each frame. Three feature extraction methods were tested, (i) conventional kinematic feature design to capture key characteristics of item 23 as described in the UPDRS manual, (ii) massive higher-order feature extraction based on the HCTSA library [2], and (iii) a deep learning neural network based on a Multi-Layer Perceptron (MLP) classifier for time series and applied to perform supervised statistical learning against reference clinical diagnosis. An ordinal classifier, based on an extra-tree classifier, was trained and evaluated using 100 random splits with a stratification strategy. Furthermore, an extra-tree classifier was trained in the subset of ON-OFF measures in the same patient, to detect changes in the UPDRS item 23 score after DRT. Classification performance is reported using a balanced accuracy score.

Samuel Camba Fdez

R&D in Computer Vision and AI | PhD Student

My research interests include artificial intelligence and computer vision. I am currently focusing my thesis on biomechanical data analysis using deep learning and weak learning techniques.