Abstract

Parkinson’s Disease (PD) is a progressive, chronic neurological disorder that is distinguished by abnormalities in the motor system. The condition can be detected in the early stage by the irregular handwriting of the individual. Early diagnosis is critical to enable timely therapeutic intervention and slow disease advancement. However, traditional diagnostic approaches largely depend on subjective clinical assessments, which lack scalability and exhibit reduced sensitivity in the prodromal phase. The present study proposes a well-established deep-learning architecture using transfer learning with MobileNetV2, which can be used for early diagnosis of Parkinson’s Disease through handwriting images. The dataset includes 816 samples from 120 people. It is augmented through grayscale and HSL to add more variety to feature samples of the model. A two-stage training regimen—initial base freezing followed by fine-tuning with a reduced learning rate—was employed to optimize convergence and generalization. The approach presented in this study score 92% on accuracy with f1-score of 0.88 and precision of 0.81 outperforming those of conventional baselines in regard to sensitivity and robustness. The resulting framework is lightweight, non-invasive, and well-suited for real-time screening applications, offering significant potential for clinical decision support and remote telehealth deployments.

Key words: Parkinson’s Disease, Handwriting analysis, Early diagnosis, Transfer learning, Deep Learning in Healthcare, Convolutional neural networks.