Towards Parkinson's Disease Detection Through Analysis of Everyday Handwriting

Background: Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder worldwide. People suffering from PD exhibit motor symptoms that affect the control of upper and lower limb movement. Among daily activities that depend on proper upper limb control is the handwriting process, which has been studied in state-of-the-art research, mainly considering non-semantic drawings like spirals, geometric figures, cursive lines, and others. Objectives: This paper analyzes the suitability of modeling the handwriting process of digits from 0 to 9 to automatically discriminate between PD patients and healthy control subjects. The main hypothesis is that modeling these numbers allows a more natural evaluation of upper limb control. Methods: Two approaches are considered: modeling of the images resulting from the strokes collected by the digital tablet and modeling of the time series yielded by the digital tablet while performing the strokes, i.e., time-dependent signals. The first approach is implemented by fine-tuning a CNN-based architecture, while the second approach is based on hand-crafted features measured upon the time series, namely pressure and kinematic measurements. Features extracted from time-dependent signals are represented following two strategies, one based on statistical functionals and the other one based on creating Gaussian Mixture Models (GMMs). Results: The experiments indicate that pressure-based features modeled with functionals are the ones that yield the highest accuracy, indicating that PD-related symptoms are better modeled with dynamic approaches than those based on images. Conclusions: The dynamic approach outperformed the image-based model, indicating that the writing process, modeled with signals collected over time, reveals motor symptoms more clearly than images resulting from handwriting. This finding is in line with previous results in the state-of-the-art research and constitutes a step forward to create more accurate and informative methods to detect and monitor PD symptoms.