Volume 16, Issue 6 And S7 (Artificial Intelligence- In Press 2026)
J Research Health 2026, 16(6 And S7): 2-2 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Pashaei M, Mortezania F, Amiri H. Unsupervised Plank Posture Assessment Using Fuzzy C-Means Clustering and Multi-Objective Threshold Optimization. , J Research Health. 2026; 16 (6 and S7) :2-2
URL: http://jrh.gmu.ac.ir/article-1-3228-en.html
URL: http://jrh.gmu.ac.ir/article-1-3228-en.html
1- Department of Engineering Sciences, College of Engineering, University of Tehran, Tehran, Iran.
2- Department of Engineering Sciences, College of Engineering, University of Tehran, Tehran, Iran. ,hadi.amiri@ut.ac.ir
2- Department of Engineering Sciences, College of Engineering, University of Tehran, Tehran, Iran. ,
Abstract: (31 Views)
Background: Correct posture during exercise is essential for maximizing physical benefits and preventing injury. Supervised machine learning models can classify posture correctness, but they depend on extensive, costly, and time-consuming expert data labelling. Simple rule-based systems (e.g., enforcing strict 90° and 180° angles) are impractical because of natural anatomical tolerances and the noise and view-dependency of pose estimation tools such as MediaPipe. This study proposes a novel unsupervised methodology to evaluate plank posture and diagnose specific anatomical errors without manually labelled datasets.
Methods: MediaPipe Pose was used to extract joint coordinates from a dataset of plank exercises, from which four key angles (elbow, shoulder, hip, and knee) were calculated. Fuzzy C-Means (FCM) clustering was applied to handle the continuous nature of human posture, with the number of clusters tuned using the Fuzzy Partition Coefficient (FPC). The cluster closest to the ideal plank angles (90-90-180-180) was identified. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) was then used for multi-objective optimization to dynamically determine the optimal membership threshold for assigning samples to the correct posture cluster, balancing cluster purity against data retention. All experiments were repeated over five random seeds. Performance was evaluated on an independent set of 69 images that were not used for clustering or threshold selection and were annotated by two raters with adjudication by a coach.
Results: FCM + NSGA-II achieved the lowest cumulative error among the FCM variants (22.17 ± 0.65°) and the highest agreement with expert annotations (κ = 0.783, F1 = 0.921), while retaining 58.2 of 69 test images. It clearly outperformed the rule-based method (κ = 0.113) and DBSCAN with matched retention, and was comparable with the other clustering methods. When all methods were evaluated on the same images that none of them discarded, the main methods performed identically (κ = 0.887, F1 = 0.966). This indicates that the benefit of FCM + NSGA-II comes from withholding ambiguous samples for expert review rather than from better clustering of easy cases
Conclusion: This study shows that data-driven thresholds can replace fixed tolerances for screening joint-angle quality. Tuning fuzzy c-means with multi-objective optimization (NSGA-II) explicitly balances how many samples are kept against how accurate they are, which single-objective clustering methods do not do. The approach is a practical and more reliable alternative to manual threshold selection and rigid rule-based criteria, and it can flag ambiguous samples for expert review. Future work should validate it on larger datasets and other movement tasks.
Methods: MediaPipe Pose was used to extract joint coordinates from a dataset of plank exercises, from which four key angles (elbow, shoulder, hip, and knee) were calculated. Fuzzy C-Means (FCM) clustering was applied to handle the continuous nature of human posture, with the number of clusters tuned using the Fuzzy Partition Coefficient (FPC). The cluster closest to the ideal plank angles (90-90-180-180) was identified. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) was then used for multi-objective optimization to dynamically determine the optimal membership threshold for assigning samples to the correct posture cluster, balancing cluster purity against data retention. All experiments were repeated over five random seeds. Performance was evaluated on an independent set of 69 images that were not used for clustering or threshold selection and were annotated by two raters with adjudication by a coach.
Results: FCM + NSGA-II achieved the lowest cumulative error among the FCM variants (22.17 ± 0.65°) and the highest agreement with expert annotations (κ = 0.783, F1 = 0.921), while retaining 58.2 of 69 test images. It clearly outperformed the rule-based method (κ = 0.113) and DBSCAN with matched retention, and was comparable with the other clustering methods. When all methods were evaluated on the same images that none of them discarded, the main methods performed identically (κ = 0.887, F1 = 0.966). This indicates that the benefit of FCM + NSGA-II comes from withholding ambiguous samples for expert review rather than from better clustering of easy cases
Conclusion: This study shows that data-driven thresholds can replace fixed tolerances for screening joint-angle quality. Tuning fuzzy c-means with multi-objective optimization (NSGA-II) explicitly balances how many samples are kept against how accurate they are, which single-objective clustering methods do not do. The approach is a practical and more reliable alternative to manual threshold selection and rigid rule-based criteria, and it can flag ambiguous samples for expert review. Future work should validate it on larger datasets and other movement tasks.
Type of Study: Orginal Article |
Subject:
● Health Education
Received: 2026/09/27 | Accepted: 2026/10/17 | Published: 2026/11/11
Received: 2026/09/27 | Accepted: 2026/10/17 | Published: 2026/11/11
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
This is an open access article distributed under the terms of the Creative Commons Attribution License (CC-By-NC), which permits use, distribution, and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes.
Contact Information
