A Deep Learning Framework for Early Parkinson’s Disease Detection: Leveraging Spiral and Wave Handwriting Tasks with EfficientNetV2-S
Ayesha Razaq, Shabana Ramzan, Sohail Jabbar, Muhammad Munwar Iqbal, Muhammad Asif Habib, Umar Raza
Abstract
Background: Early detection of Parkinson’s disease (PD) is vital for improving patient outcomes, yet traditional diagnostic methods often depend on subjective clinical evaluations. Methods: This study proposes a novel deep learning framework for PD detection based on spiral and wave handwriting patterns from the PaHaW dataset. A comprehensive preprocessing pipeline is implemented, integrating histogram equalization and Canny edge detection. The processed spiral and wave images are evaluated independently using a fine-tuned EfficientNetV2-S architecture for binary classification. In addition to the EfficientNetV2-S experiments, a baseline Convolutional Neural Network (CNN) model is implemented separately for the spiral and wave handwriting images. The proposed model is further assessed using a 5-fold cross-validation strategy to ensure robustness and generalizability. Results: The models achieved validation accuracies of 98.68% on the spiral dataset and 98.10% on the wave dataset, with high Receiver Operating Characteristic–Area Under the Curve (ROC–AUC) scores, indicating robust discrimination between healthy and PD subjects. Analysis of the confusion matrix and classification results confirmed consistent sensitivity and specificity across the dataset. The 5-fold cross-validation yielded a standard deviation of ±0.0109. Conclusions: These results highlight the strong potential of handwriting analysis for early PD detection.