Kinetic Pattern Recognition in Home-Based Knee Rehabilitation Using Machine Learning Clustering Methods on the Slider...

Recent advancements in rehabilitation sciences have progressively used computational techniques to improve diagnostic and treatment approaches. However, the analysis of high-dimensional, time-dependent data continues to pose a significant problem. Prior research has used clustering techniques on rehabilitation data to identify movement patterns and forecast recovery outcomes. Nonetheless, these initiatives have not yet used force or motion datasets obtained outside a clinical setting, thereby limiting the capacity for therapeutic decisions. Biomechanical data analysis has demonstrated considerable potential in bridging these gaps and improving clinical decision-making in rehabilitation settings.

This study presents a comprehensive clustering analysis of multidimensional movement datasets captured using a novel home exercise device, the "Slider". The aim is to identify clinically relevant movement patterns and provide answers to open research questions for the first time to inform personalized rehabilitation protocols, predict individual recovery trajectories, and assess the risks of potential postoperative complications.

High-dimensional, time-dependent, bilateral knee kinetic datasets were independently analyzed from 32 participants using four unsupervised clustering techniques: k-means, hierarchical clustering, partition around medoids, and CLARA (Clustering Large Applications). The data comprised force, laser-measured distance, and optical tracker coordinates from lower limb activities. The optimal clusters identified through the unsupervised clustering methods were further evaluated and compared using silhouette analysis to quantify their performance. Key determinants of cluster membership were assessed, including demographic factors (eg, gender, BMI, and age) and pain levels, by using a logistic regression model with analysis of covariance adjustment.

Three distinct, time-varying movement patterns or clusters were identified for each knee. Hierarchical clustering performed best for the right knee datasets (with an average silhouette score of 0.637), while CLARA was the most effective for the left knee datasets (with an average silhouette score of 0.598). Key predictors of the movement cluster membership were discovered for both knees. BMI was the most influential determinant of cluster membership for the right knee, where higher BMI decreased the odds of cluster-2 membership (odds ratio [OR] 0.95, 95% CI 0.94-0.96; P<.001) but increased the odds for cluster-3 assignment relative to cluster 1 (OR 1.05, 95% CI 1.03-1.06; P<.001). For the left knee, all predictors of cluster-2 membership were significant (.001≤P≤.008), whereas only BMI (P=.81) could not predict the likelihood of an individual belonging to cluster 3 compared to cluster 1. Gender was the strongest determinant for the left knee, with male participants significantly likely to belong to cluster 3 (OR 3.52, 95% CI 2.91-4.27; P<.001).

These kinetic patterns offer significant insights for creating personalized rehabilitation procedures, potentially improving patient outcomes. These findings underscore the efficacy of unsupervised clustering techniques in the analysis of biomechanical data for clinical rehabilitation applications.