XAI-OPL Risk Prediction Dataset — 31 Studies on AI/ML for Occupational Postural and Lumbar Risk

Abstract

This dataset is the complete reproducibility package accompanying the systematic review and meta-analysis "Explainable and interpretable machine learning for occupational postural and lumbar risk prediction: A systematic review and meta-analysis" (PROSPERO registration: CRD420261530250). The review is, to our knowledge, the first to explicitly evaluate the role of explainable artificial intelligence (XAI) in AI/ML-based occupational postural and lumbar risk prediction. The package contains the full raw data extraction matrix for all 31 included studies, the meta-analysis input data and outputs (pooled accuracy, pooled AUC, subgroup analyses by model type, sector-level performance, and publication bias tests), the PROBAST risk-of-bias assessments, and Python scripts that regenerate every figure and table reported in the manuscript. The dataset is designed to support full analytical reproducibility — any researcher can reproduce the manuscript's results by re-running the scripts against the provided CSVs, without requiring access to the original study data. The 31 included studies were published between 2017 and 2026, span six occupational sectors (Industrial/Manufacturing, Construction/Logistics, Healthcare, Office/Telework, Agriculture, and Mixed/Multi-sector), and employ five data modalities (computer vision, wearable inertial measurement units, surface electromyography, 3D keypoints, and multimodal sensors). The extracted fields cover 11 thematic blocks: bibliographic identifiers, population and occupational context, exposure and task characteristics, AI/ML model architecture, data modality and input features, prediction task and outcome, validation strategy, performance metrics, explainability and interpretability methods, PROBAST risk-of-bias domains, and key biomechanical predictors. Values not reported in the original publications are marked NR (Not Reported); values not applicable (e.g., for non-AI/ML studies) are marked NA (Not Applicable). The dataset is released under a CC BY 4.0 license to maximise reuse in evidence synthesis, methodological benchmarking, and educational contexts.