The role of signal preprocessing on the discriminability of Canonical Time-series Characteristics and classification among individuals with and without Parkinson's disease during Serious Game interaction

Abstract

Over the past decade, there has been a significant increase in studies using biomedical signals for objective monitoring of Parkinson's disease (PD) motor symptoms. Inertial sensors are widely employed to record motion, producing time-series data that capture the underlying motor condition of patients. A major challenge in the field is classifying these signals to discriminate healthy subjects from PD individuals and distinguish motor conditions among patients. While many studies focus on feature classification, there is a lack of research on the influence of signal preprocessing. To fill this gap, we evaluate data from healthy subjects and PD patients during interaction with the RehaBEElitation serious game. We employed the catch22 feature set to extract robust time-series characteristics. To evaluate the influence of preprocessing on classification between healthy individuals and patients in "on" and "off" medication states, four strategies were adopted. Initially, features extracted from raw data showed limited accuracy due to noise and voluntary movements. Subsequent interpolation to address discontinuities produced inconsistent results. The third strategy involved wavelet decomposition, which effectively mitigated trends and motion artifacts, resulting in a significant increase in accuracy across all models and confirming the vital role of sophisticated signal filtering. The fourth strategy combined interpolation and wavelet Signal Processing and PD Classification decomposition, achieving the best results with optimal separation (Accuracy = 100.0%) in binary classification and significant improvement in the multi-class problem. Our findings establish that signal conditioning is pivotal for maximizing discriminative power. To further validate our findings, we benchmarked our pipeline against the RandOm Convolutional KErnel Transform (ROCKET) using a RidgeClassifierCV. The catch22-RF, using wavelet-based approach, achieved a balanced accuracy of 76.0% in the multi-class task, demonstrating superior performance compared to the ROCKET-RidgeClassifierCV framework (69.0%) while maintaining a more compact and computationally efficient feature representation.