Parkinson’s Disease Early Detection using Hand-drawn Data with Deep Learning, Ensemble Learning, and Explainable AI

Abstract

Parkinson's Disease (PD) impairs the brain's ability to control movements. Early diagnosis remains challenging due to limitations in conventional diagnostic methods, which are expensive and technically demanding, and gait and voice analysis, which are non-invasive but may require specialized equipment or controlled environments. This research aims to develop and evaluate a cost- effective, non-invasive, and accurate method for early PD diagnosis using hand-drawn images using deep learning models, the Ensemble method, and Explainable AI. A dataset of 3,264 images (1,632 healthy and 1,632 PD) was normalized, augmented (using flipping, zooming, and rotation), and split into training, validation, and test sets. Pre-trained models like VGG16, VGG19, ResNet50, and DenseNet121 were used for feature extraction, followed by custom classification layers for final prediction. Among individual models, DenseNet121 achieved the best results, with 98\% accuracy, 0.97 sensitivity, and 0.99 specificity. The soft-voting ensemble (VGG16, ResNet50, and DenseNet121) outperformed it, attaining 99\% accuracy, 0.98 sensitivity, and 1.00 specificity. Employed XAI techniques such as Grad-CAM, LIME, and SHAP to enhance interpretability, with Grad-CAM providing the most effective visual explanations. Although this approach requires moderate computational resources, it establishes a foundation for future multimodal diagnostic systems integrating EEG, MRI, and voice data to improve diagnostic confidence further.