On the Application of DeepLabCut for the Assessment of Spiral and Sinusoidal Patterns in Individuals with Parkinson’s Disease

Abstract

Early diagnosis of Parkinson’s disease (PD) is a clinical challenge due to the non-specific nature of the initial motor symptoms (MS) and the limitations of traditional diagnostic methods. Hand-drawn tasks, such as Archimedes’ spiral, have emerged as potential tools for assessing motor impairment. This study aimed to employ computer vision and machine learning (ML) techniques, with an emphasis on the use of DeepLabCut (DLC), to analyze drawing patterns of individuals with PD and healthy controls. Twenty individuals (10 with PD and 10 healthy controls), matched for age and gender, participated in the study. Each participant completed 12 manual drawings, totaling 240 samples recorded on video. The recordings were processed in DLC to extract spatial coordinates, and the resulting signals were preprocessed and normalized. Eighty-five variables related to amplitude, frequency, entropy, statistical properties, and displacement were extracted. For classification, attribute selection techniques and 14 supervised ML algorithms were applied, using cross-validation (CV) for performance evaluation. The ML models applied to the drawing patterns demonstrated high classification ability, with an accuracy of approximately 93%. The Extra Trees Classifier (ETC) showed the highest performance, reaching an AUC of 0.98, followed by K-Neighbors with an AUC of 0.97. Attribute selection contributed to performance optimization, and CV strategies ensured the robustness of the results. These findings highlight the potential of DLC and ML as a non-invasive tool for early detection and monitoring of motor impairment in Parkinson’s disease.