Enhanced Early Parkinson’s Classification: A Transfer Learning Approach Using Augmented Hand Drawing Records

Abstract Parkinson's disease (PD) is a brain disorder affecting motor control due to damage in the basal ganglia, leading to the death of about 60% of neurons and a significant drop in dopamine. This results in motor impairments like tremors and makes movement and communication harder as the illness advances. Micrographia, small or atypical handwriting, can be an early sign and to address early diagnosis, this study introduces a novel Machine Learning (ML) approach. The proposed workflow uses transfer learning with the EfficientNetB0 model, employing learning curve analysis across network metrics to optimize depth, width, and resolution. Experimental results show this method effectively predicts early PD with high metrics, 98.77% accuracy, 99.07% precision, 99% ROC, 96.94% F1 score, and 98.47% recall and specificity. This enhanced performance is due to the EfficientNetB0 model's architecture having fewer parameters and FLOPS, utilizes MBConv blocks, depth-wise separable convolutions, and ReLU activation. The study uses a dataset of original and augmented spiral and wave drawings to analyze PD-related motor impairments. The high prediction accuracy of this EfficientB0-based approach enables early-stage classification, which is vital for grouping patients with similar traits to better understand and treat PD.