Integration of Kinematic Features with Micropause, Rotation Consistency, and DTW for Parkinson’s Disease Classification

Handwriting-based assessment has emerged as a promising digital biomarker for Parkinson's Disease (PD), yet the generalizability of existing models across different datasets remains unclear.This study investigates spiraldrawing data collected via a smartphone-stylus system and proposes a feature-based classification approach integrating kinematic measures, 1D magnitude Dynamic Time Warping, rotation consistency, and micropause descriptors.A Random Forest classifier was evaluated using internal cross-validation and external cross-dataset testing on two public benchmarks: ParkinsonHW and PaHaW.On the internal dataset, the proposed feature combination achieved a high accuracy of 99.23%, indicating strong discriminative power within-domain.However, performance varied substantially when deployed on external datasets.Accuracy reached 93.42% on ParkinsonHW but dropped to 65.33% on PaHaW, suggesting dataset-dependent generalization.Comparative analysis with prior studies showed that methods leveraging deep temporal models such as LSTM-CNN achieved higher accuracy on PaHaW, highlighting the limitations of handcrafted features and 1D temporal alignment.Domain adaptation using CORAL improved performance over normalization alone but did not fully bridge cross-dataset gaps.Overall, the findings emphasize the potential of lightweight feature-based models while underscoring the need for feature-device invariance, richer temporal representations, and harmonized data collection protocols for robust PD handwriting assessment.