Clustering Kinematic Patterns of Runners: A Comparative Study of Hierarchical, K-Means, and Deep Temporal Clustering Algorithms

• Published in: Faṣlnāmah-i ʻilmī-pizhūhishī-i Vol. 14; no. 5; pp. 698 - 709
• Main Authors: Mohamadi, Zaniar, Eslami, Mansour, Yousefpour, Rohollah
• Format: Journal Article
• Language: English
• Published: 01.11.2025
• ISSN: 2251-8401; 2252-0414; 2252-0414
• DOI: 10.32598/SJRM.14.5.3363

Background and Aims Current biomechanical studies lack a comprehensive investigation of deep learning clustering algorithms for identifying homogeneous movement patterns in runners. This study aims to compare the performance of principal component analysis (PCA)-based hierarchical clustering and K-means algorithms with an end-to-end deep temporal clustering (DTC) approach in analyzing ankle joint kinematics of runners with homogeneous movement patterns. Methods Three-dimensional ankle joint angles were obtained from 108 recreational runners (55 males and 53 females; age: 22.45±2.42 years, height: 1.69±0.11 m, body mass: 64.64±9.54 kg) during barefoot running at a speed of 3.0±3 meters per second. DTC, hierarchical, and K-means algorithms were trained using ankle joint angles during running. After clustering, the performance and accuracy of each algorithm in identifying clusters with homogeneous movement patterns were evaluated by calculating the Silhouette score, the Calinski-Harabasz index (CHI), and the Davies-Bouldin index (DBI). Results In cluster separation, the DTC algorithm demonstrated superior performance and accuracy compared to the other two algorithms (silhouette score=0.74, DBI=0.35). This algorithm identified three distinct clusters with a clustering inconsistency rate of 6%. The hierarchical clustering method achieved a silhouette score of 0.68 and a DBI value of 0.52 in 10 seconds with a 15% inconsistency rate. The K-means method showed a silhouette score of 0.63 and a DBI of 0.78 in 3 seconds with an 18% inconsistency rate. Conclusion The DTC algorithm outperforms hierarchical clustering and K-means clustering in identifying homogeneous movement patterns among runners. Its higher accuracy and lower learning error make it a suitable choice for analyzing kinematic data in biomechanical research. The findings can enhance the understanding and analysis of movement patterns and contribute to the development of effective strategies for prescribing targeted interventions.