Governance-Aware Machine Learning for Quality Event Risk Flagging in Regulated Laboratory Operations

This study examines the potential of machine learning to support quality-event risk flagging in regulated laboratory operations. A domain-informed synthetic dataset comprising 500 observations was used to compare logistic regression, random forest, and XGBoost models. The evaluation considered predictive discrimination, precision–recall performance, probability calibration, robustness to controlled data-quality perturbations, sanity checks, and permutation importance. Random forest achieved a ROC-AUC of 0.963 and a PR-AUC of 0.957 in the synthetic benchmark. Calibration-related variables, particularly time since last calibration and calibration interval, made the largest contributions to model performance. The study provides an applied machine-learning benchmarking framework for investigating calibration review and preventive-maintenance prioritization. Its findings are limited to synthetic data and do not establish real-world predictive validity or operational readiness. Further validation with real laboratory records is required. Publication status: This manuscript is a preprint and has not undergone formal journal peer review.