Machine learning applied for the motor pattern classification of pen-and-paper spiral drawings in Parkinson’s disease

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by impaired voluntary motor control, most commonly manifesting as tremors, bradykinesia, and muscle rigidity. Despite significant advancements in the understanding of PD, clinical evaluation remains the primary method for diagnosis and disease monitoring, often relying on subjective assessments. Spiral analysis has emerged as a promising, low-cost, and noninvasive technique for the objective quantification of motor symptoms through the study of hand-drawn spirals. Unlike previous studies that utilized digitizing tablets and digital pens, the present research investigates the feasibility of using traditional pen-and-paper spirals in a clinical setting. This study aims to develop objective tools to assist healthcare professionals in clinical decision-making. It is part of a broader project called PARKINSON BRASIL, which seeks to develop a comprehensive national database containing multidimensional aspects of individuals with Parkinson’s disease. Four deep learning architectures— VGG16, ResNet50, MobileNetV2, and Xception—were employed for feature extraction, followed by classification using error-correcting output codes (ECOC). Participants were classified into either PD or control groups, and PD participants were further categorized into disease stages according to the Hoehn and Yahr scale. Model performance was evaluated using standard metrics, including accuracy and F1-score. Xception achieved the best overall performance. ResNet50 and VGG16 also demonstrated high precision in their performance.The findings demonstrate that pen-and-paper spiral analysis, when integrated with machine learning techniques, holds significant potential as a complementary tool for the diagnosis and monitoring of Parkinson’s disease in clinical environments.