ParkXplainer: An Interpretable Ensemble Learning Driven Approach for Hand-drawn Image based Early Parkinson's Detection using Grad-CAM

Gayatri SaxenaSeema BawaRohit Ahuja^*

• Thapar Institute of Engineering and Technology (Deemed to be University), Indra Puri Colony, India

Parkinson's Disease (PD) affects motor control, making early detection difficult through traditional approaches. This research proposes a cost-effective, deep learning–based, non-invasive, diagnostic method using hand-drawn spiral and wave images. Two datasets were used: a primary dataset of 3,264 images (1,632 PD and 1,632 Healthy) and a secondary dataset of 204 images. The models were trained and validated using stratified sampling and K-Fold cross-validation on the primary dataset, while K-Fold and Two-Cross-Validation techniques were applied to the secondary dataset. Pre-trained CNN architectures (VGG16, VGG19, ResNet50, and DenseNet121) were utilized for feature extraction and classification. Among these, the Equal-Weighted Soft Voting Ensemble achieved the best performance on the primary dataset, with 99% accuracy, 0.98 sensitivity, and 1.00 specificity, demonstrating strong generalization with minimal overfitting. Explainable AI techniques (Grad-CAM, LIME, and SHAP) further enhanced model interpretability, with Grad-CAM effectively highlighting affected regions in PD drawings. The proposed ParkX-plainer method offers a low-cost screening tool deployable via tablets/mobiles, supporting early neurological consultation.