Meta-analysis of AI posture recognition for work-related musculoskeletal disorders prevention in occupational...

The objective of this systematic review and meta-analysis was to evaluate the performance of artificial intelligence (AI) posture recognition systems in preventing the occurrence of work-related musculoskeletal disorders (WMSDs) across all occupational activities, following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines.

During May 2026, six databases were screened without publication date restrictions. Selected studies had to evaluate the performance of a posture recognition system using a machine learning (ML) or deep learning (DL) approach to assess the risk level of WMSD development using at least one of the following metrics: accuracy, specificity, sensitivity, precision, and F1 score. Risk of bias, heterogeneity, sensitivity, publication bias, and certainty of evidence analyses were conducted in the meta-analysis. Logit and Freeman-Tukey transformations were applied to the data before pooling and results were compared to those obtained from the untransformed. A meta-analysis was performed by performance parameter and by transformation. Four subgroup analyses were performed: comparison of AI methods, sensor types, professional activities, and ergonomic tools used, using pairwise mixed-effects meta-regressions.

Among the 1,071 unique entries, 26 studies were included. The logit transformation most closely approximated a normal distribution for all performance metrics without affecting heterogeneity, which was high (I² > 90%) in all analyses. Overall performance ranged from 85 to 95% for the five performance metrics. DL significantly outperformed ML by 4 to 16% for specificity, sensitivity, and accuracy, regardless of the transformation (p ≤ 0.05). Hybrid solutions demonstrated higher sensitivity and accuracy (> 99%, p < 0.05) with low heterogeneity (I² < 30%) regardless of the transformation. Occupational activity had a limited influence on overall performance, and no assessment tool consistently outperformed all others across all five performance parameters.

In conclusion, the logit transformation consistently improved the statistical properties of the data and AI approach, type of sensors, occupational activity, and ergonomic risk assessment tool all significantly influenced the reported performance metrics. However, findings should be interpreted with caution and prudence is required regarding the deployment of the systems under real-world conditions due to the very low level of evidence.

https://www.crd.york.ac.uk/PROSPERO/home, PROSPERO, n°: CRD420261432405.