AI-based models for risks of work-related musculoskeletal disorders in office workers: exploring generalisability, interpretability and use of binary scales.
TL;DR
Previously developed machine learning models were externally validated and the performance of simplified binary data inputs evaluated, showing the feasibility of simplified, interpretable, and scalable AI-based tools for practical ergonomic risk assessment in office environments.
Abstract
Artificial intelligence has the potential to enhance risk assessment of work-related musculoskeletal disorders in office environments. This study externally validated previously developed machine learning models and evaluated the performance of simplified binary data inputs. Six models trained on data from 810 office workers were validated using an independent dataset of 74 globally recruited participants. Model performance across nine anatomical regions was assessed using F1-scores and AUC-ROC metrics. SHapley Additive exPlanations were used to interpret risk factor contributions, and model performance using ordinal Likert-scale inputs was compared with binary-transformed representations. Generalisability varied by anatomical region, with moderate performance for high-prevalence regions (neck and lower back) and reduced accuracy for low-prevalence regions. XGB showed moderate but comparatively stable performance across several anatomical regions. These findings highlight the need for region-specific validation and support the feasibility of simplified, interpretable, and scalable AI-based tools for practical ergonomic risk assessment in office environments.