Integrating Handwriting Modalities Through a Lightweight Hybrid Transfer Learning Model for Early Parkinson’s Disease Diagnosis

Abstract

Background: Since the initial motor symptoms are subtle and progress gradually, it is challenging to diagnose PD in the early stages. In this study, we present a flexible hybrid transfer learning framework to early detect PD using heterogeneous handwriting patterns, which combines offline and online handwriting data to characterize PD-related motor dysfunctions, and uses pre-trained deep learning to enhance early detection.

Methods: Spiral drawing images are used to identify spatial and structural features, and timing, pressure, and movement information from online handwriting data in the PaHaw dataset. Data quality is enhanced using a hybrid preprocessing pipeline, such as handling missing values, class balancing, and Min–Max normalization, and a feature selection technique is applied that combines parallel and sequential techniques to reduce redundancy and increase the relevance of features.

Results: The model showed nearly perfect classification accuracy of 99.99 percent, confirming its generalization and discriminative capability, which suggests that the framework is a highly accurate, non-invasive and effective method for early detection of PD and that it can be used in a clinical setting in real-world conditions, even in resource-limited settings.