STAMP-Informed Machine Learning for Process Safety Event Prediction: Synthetic Dataset and Reproducibility Material

Logistic regression using lagging indicators only. 2. Logistic regression using STAMP-informed leading indicators. 3. Gradient boosting using the same STAMP-informed leading indicators.

The analysis incorporates temporally separated model development, validation, and testing, alongside discrimination, calibration, precision–recall analysis, cluster-bootstrap comparisons, and permutation-based feature importance. The accompanying supplementary materials include the synthetic dataset, Python analysis script, model performance results, paired model comparisons, test-set predictions, calibration data, feature-importance results, and supporting visualizations.

Data provenance and limitations: all observations, operational indicators, and process safety events are computationally simulated. No actual industrial incident records, confidential operational data, or personal information are included. The dataset is intended for methodological demonstration, computational reproducibility, and further research.

Its results do not establish real-world predictive performance or validate the models for operational safety-critical decision-making. The materials should be interpreted alongside the associated research manuscript. Application to actual industrial operations would require site-specific hazard analysis, validation using real operational data, and appropriate safety assurance.