Evaluating Circle-Tracing Performance Using Machine Learning in a Haptic Interaction Framework.

Hand and wrist function is commonly assessed using qualitative clinical tests that provide limited objective information and data in motor performance over time. Advances in haptic technology and machine learning provide new opportunities to develop more quantitative approaches. In this study, we developed a quantitative circle-tracing assessment using a haptic pen. The primary objective was to use unsupervised machine learning to compare low performance and high-performance repetitions to determine if there were significant differences in variability. Twenty-nine healthy participants completed 40 circle-tracing repetitions while positional data were recorded and converted into tracing error metrics. Circle drawing was performed with constrained upper arm and wrist. Tracing performance was analyzed using sliding-window percentage-change calculations and a Bayesian Gaussian Mixture Model (BGMM) built into K-fold Cross Validation to identify stable clusters representing high-performance, low-performance, and neutral states. The proposed framework successfully identified significant performance fluctuations across multiple time intervals. High-performance trials were detected at 90% probability, whereas low-performance trials were identified at a greater probability (98%). Comparing the two cases, low-performance detection demonstrated less variability across short- and long-term evaluation windows, with higher agreement across multiple window sizes. In contrast, high-performance detection exhibited greater variability. Both high and low-performance detections were most reliable (97.5% and 99.7%, respectively) during short-term assessments. These findings suggest that declines in performance are more readily distinguishable than performance improvements when using error-based circle-tracing metrics. The results demonstrate the feasibility of combining haptic assessment with unsupervised machine learning to objectively characterize circle-tracing performance. This approach has potential to improve diagnosis and rehabilitation monitoring by providing quantitative measures.