AI Posture Recognition Performance for Work-Related Musculoskeletal Disorders Prevention in Manufacturing: Comparison Between Logit and Freeman-Tukey Transformation in Meta-Analysis
TL;DR
High heterogeneity (I2 > 90%) and substantial inter-study variability were observed in all analyses, and a very low level of evidence was evidenced for all performance parameters, but the results should be interpreted with caution, particularly regarding the deployment of the systems in real-world settings.
Abstract
The objective of this study was to assess the performance of posture recognition systems based on artificial intelligence (AI) using deep learning (DL) and machine learning (ML) approaches for the prevention of work-related musculoskeletal disorders (WMSDs) in manufacturing. The study was conducted as a systematic review and meta-analysis following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Four open-access databases were screened in May 2026 without date restrictions: PubMed/MedLine, Google Scholar, ScienceDirect, and IEEE Xplore. The selected studies had to be original, peer-reviewed studies written in English. The study had to evaluate the performance of an AI posture recognition system (ML or DL) for the prevention of WMSDs in manufacturing using at least one of the following parameters: accuracy, specificity, sensitivity, precision, or F1 score. The risk of bias for each included study was assessed using PROBAST (Prediction Model Study Risk of Bias Assessment Tool). A meta-analysis was conducted to pool the values of the five performance metrics separately. Logit and Freeman-Tukey transformations were applied prior to pooling, and the results were compared after back-transformation. Cochran’s Q test, the I2 statistic, and inter-study variability (τ2), computed using the generalized inverse variance method with the restricted maximum likelihood model, were applied to assess heterogeneity. Forest plots including pooled values with 95% confidence intervals were used to present the results. Subgroup analyses and meta-regressions were performed to test the effect of AI methods and ergonomic tools on performance and to explore potential causes of heterogeneity. Finally, publication bias (Egger’s test) and certainty of evidence (GRADE method—Grading of Recommendations Assessment, Development, and Evaluation) were assessed to ensure the generalizability of the results. Ten studies were included: Among the 200 studies identified through database searches and the snowball method, 12 met the inclusion criteria and were selected. Two studies were excluded due to an insufficient number of participants, bringing the final number of studies considered to 10. The logit transformation yielded the best overall fit for the normality of the data distribution for the use of a random-effects model. High posture detection performance was observed, with pooled values ranging from 84.78% (95% CI: 80.23–88.19%, specificity with Freeman-Tukey) to 93.40% (95% CI: 89.57–95.89%, precision with logit). The values obtained with logit were higher than those obtained with Freeman-Tukey, with differences ranging from 2.75% to 4.75% across all performance metrics. Meta-regression showed that DL outperformed ML for all metrics, with differences ranging from 5% to 17%. RULA and REBA achieved better performance than other ergonomic tools. However, high heterogeneity (I2 > 90%) and substantial inter-study variability were observed in all analyses, and a very low level of evidence was evidenced for all performance parameters. Consequently, the results should be interpreted with caution, particularly regarding the deployment of the systems in real-world settings. Future work could strengthen training and testing procedures on datasets, as well as external validation. These aspects are essential for effective use in manufacturing environments to ensure operator safety by reducing their exposure to WMSD risks associated with work postures.