Prediction of musculoskeletal discomfort through integration of item response theory and machine learning

This study examined the feasibility of integrating Item Response Theory (IRT) and Machine Learning (ML) algorithms to predict musculoskeletal discomfort levels among workers across various sectors. IRT provided a psychometrically robust, continuous measure of discomfort from self-reported symptoms, while ML enabled automated and scalable classification. This combined approach overcomes limitations of traditional sum-score methods and static analyses, offering a novel pathway for dynamic ergonomic assessment. Symptom data from 300 workers across 25 body regions were analyzed using the IRT graded-response model to estimate latent trait scores. Based on these scores, discomfort levels were defined and used as response variables in supervised ML models. The tested algorithms included logistic regression, Support Vector Machine (SVM), K-nearest neighbors (KNN), decision trees, Random Forest, and eXtreme Gradient Boosting (XGBoost), evaluated using stratified cross-validation. To improve class balance, the Synthetic Minority Over-sampling Technique Nominal Continuous (SMOTE-NC) algorithm was applied. Performance evaluation employed precision, accuracy, recall, and F-score, with the SVM and KNN models producing the best results. Analysis using Shapley Additive exPlanations (SHAP) identified symptoms in the forearm, hip, and feet as the main predictors of higher discomfort levels. The proposal is innovative in combining the psychometric sensitivity of IRT with the predictive capacity of ML algorithms, paving the way for future digital solutions for real-time ergonomic screening and monitoring. This work helps fill a gap in the literature by proposing a replicable and accurate approach for automated assessment of occupational musculoskeletal discomfort, with potential for scalability in future applications. • IRT and ML integration enables automated prediction of discomfort in workers. • SVM and KNN models achieved the best predictive performance for discomfort levels. • SHAP analysis highlights feet, hips, and forearms as key predictive regions. • SMOTE-NC improved model accuracy by balancing discomfort class distribution. • Method supports future digital tools for real-time risk screening and prevention.