An Investigative Study of Widely Used Ergonomic Risk Assessment Methods Based on Machine Learning and MCDM Approaches

OpenAlex İngilizce

REBA, RULA, and OWAS are widely used observational ergonomic risk assessment methods based on predefined ordinal scoring rules. The final risk scores generated by these methods play a critical role in determining and prioritizing ergonomic interventions under limited resources to reduce musculoskeletal risks and related losses. Because the corresponding scores are deterministic functions of their component inputs, the present study focuses on machine-learning (ML) models as surrogate-analysis tools rather than replacements for exact scoring procedures. Ordinal Logistic Regression (AT, IT, SE), Random Forest (RF), Ordinal XGBoost, and Ordinal LightGBM were evaluated on full-factorial rule-space datasets using 5 × 3 nested cross-validation, 95% confidence intervals, and Friedman tests with Nemenyi post-hoc comparisons. Local transition fidelity and controlled adjacent-category and Gaussian perturbations were additionally examined against the exact scoring rules, while Random-Forest SHAP analysis was used to characterize the fitted surrogate structures. RF provided the highest overall clean-rule-space fidelity for REBA (MAE = 0.287, QWK = 0.982), RULA (MAE = 0.011, QWK = 0.991), and OWAS (MAE = 0.244, QWK = 0.929). However, under controlled input perturbations, the exact scoring procedures generally retained lower error than the clean-trained RF surrogates, indicating that the results do not support replacing the original rules with ML. Weight-sensitivity analysis further showed that the SRP-based method ranking was conditional on criterion weights: RULA ranked first in 52.23% of 100,000 sampled weight vectors, followed by OWAS (37.43%) and REBA (10.35%). Overall, the framework provides a reproducible way to examine surrogate fidelity, local rule-space transitions, model interpretation, and the weight sensitivity of ergonomic method selection.