A Clinically-Guided Machine Learning Framework for Operational Health Risk Tier Forecasting in Construction Workers Using Wearable Data

Abstract

Real-time physiological monitoring offers a promising tool for proactive safety management in high-risk construction environments, yet its practical use is hindered by the lack of reliable clinical outcome labels and strong inter-individual variability. This study proposes a weakly supervised health-risk forecasting framework that integrates clinical-style physiological scoring, analytic hierarchy process (AHP) weighting, unsupervised clustering, and supervised learning to enable early prediction of operational risk tiers. A total of 42 627 de-identified wristband measurements from 24 construction workers were analyzed, including heart rate, body temperature, systolic and diastolic blood pressure, and oxygen saturation. Composite risk indices were generated using guideline-informed scoring and AHP weighting and grouped into four risk tiers (Low, Medium, High, Extreme) via K-means clustering to serve as proxy outcome labels. XGBoost, Random Forest, and Logistic Regression models were evaluated using strict leave-one-worker-out cross-validation. Across unseen workers, the proposed framework achieved stable discrimination of Extreme-risk states, with recall approaching 0.95 and AUC exceeding 0.97. Bootstrap analysis further confirmed the robustness of Extreme-risk detection under irregular sampling and class imbalance. These results indicate the feasibility of reliable early risk warning using wearable physiological data for construction safety management.