- The paper introduces a new database of handwriting samples from Parkinson's disease (PD) patients and healthy controls to evaluate handwriting features as potential biomarkers for modeling PD.
- Utilizing kinematic, neuromotor, and non-linear dynamics features, the study achieved classification accuracies up to 97.2% in distinguishing PD from control groups using SVM, KNN, and MLP classifiers.
- The findings indicate that online handwriting analysis shows promise as a non-invasive biomarker for PD diagnosis, highlighting the importance of using both young and elderly controls to differentiate disease effects from age-related changes.
The paper "Characterization of the Handwriting Skills as a Biomarker for Parkinson's Disease" introduces a new database of handwriting samples from Parkinson's Disease (PD) patients and healthy controls (HC). The goal of the paper is to evaluate handwriting patterns as potential biomarkers for modeling PD. The paper leverages kinematic, neuromotor and non-linear dynamic features, extracted from a database of 935 handwriting tasks collected from 55 PD patients and 94 HC subjects. The HC group was composed of both young and elderly subjects to differentiate between patterns associated to PD and patterns associated to the natural degradation of neuromotor abilities with age. The authors employed SVM (Support Vector Machines), KNN (k-Nearest Neighbors), and MLP (Multilayer Perceptron) classifiers to discriminate between PD and HC subjects.
The methodology involved the following key steps:
- Data Acquisition: Handwriting signals including x-position, y-position and pressure were recorded using a Wacom Cintiq tablet at 180 Hz. Participants performed 17 different handwriting tasks, including writing letters, digits, their ID, name and signature, a free sentence, the alphabet, and geometrical figures.
- Feature Extraction: From the recorded signals, the authors extracted kinematic features, non-linear dynamics features, and neuromotor features. Eleven statistical functionals, such as mean, median, standard deviation, and percentiles, were computed from these features.
- Kinematic features included mean velocity, max acceleration and total duration.
- Nonlinear dynamics features were computed to model stability and non-stationarity in muscular movements. These included correlation dimension, Lempel-Ziv complexity, largest Lyapunov exponent, Hurst exponent, empirical mode decomposition, and entropy.
- Neuromotor features were based on the Sigma-Lognormal model, which decomposes the velocity profile of handwriting into stroke velocity signals.
- Classification and Parameter Optimization: The classifiers used were KNN, RBF-SVM (Radial Basis Function Support Vector Machine) and MLP. The meta-parameters of the classifiers were optimized using a leave-one-out cross-validation strategy in a grid-search.
Key results from the classification experiments include:
- Classification accuracies between 81% and 97% in distinguishing PD patients from HC subjects.
- The alphabet and signature tasks yielded the best classification accuracy, with over 90% for YHC (Young Healthy Controls) vs PD and over 70% for EHC (Elderly Healthy Controls) vs PD.
- Combining all tasks in a late-fusion strategy improved the results compared to individual models. Specifically, the combination of all features with SVM yielded 96.9% accuracy for YHC vs PD, 81.7% for EHC vs PD, and 97.2% for YHC vs EHC.
The authors conclude that online handwriting analysis shows potential as a biomarker for PD diagnosis. The paper highlights the importance of using both young and elderly control subjects to differentiate between motor degradation caused by the disease and age-specific effects. They propose future work including a taxonomic paper of each task based on medical specialists' experience, analysis of offline writing samples, expansion of the database for disease progression analysis, and exploration of other classification and feature extraction techniques.