Supplemental Materials and Reproducibility Code for "From Reactive Records to Proactive Indicators: Explainable Near-Miss Diagnostics in Construction"

Supplemental Materials and Reproducibility CodeManuscript: From Reactive Records to Proactive Indicators: Explainable Near-Miss Diagnostics in ConstructionJournal: ASCE Journal of Management in Engineering (under review)Author: Joosung Lee, Department of Urban Architecture, Kangwon National University, Samcheok-si, Gangwon-do, Republic of KoreaORCID: 0000-0001-5509-9368This repository hosts the supplemental materials and the reproducibility code for the above manuscript. ASCE no longer hosts supplemental files with the article (policy effective Jan 5, 2025); this deposit is the externally hosted location referenced in the body of the manuscript.---Contents```.├── Supplemental_Materials.docx # Supplemental document (Tables S1–S8, Fig. S1)└── Python_Code/ # Reproducibility code package ├── README.md # How to run; environment; expected outputs ├── requirements.txt ├── prepare_data.py # Excel → analysis CSV (RAW progress rate; KOSHA 24→8) ├── run_analysis.py # End-to-end pipeline entry point ├── smoke_test.py # Quick self-check ├── src/ # data_preprocessing, model_training, shap_analysis, │ # eval_robustness, data_loader ├── configs/ # config.yaml + Y1/Y2/Y3 and KOSHA mapping dictionaries └── docs/ # REPRODUCIBILITY_NOTES.md, DATA_SCHEMA.md```Supplemental document (`Supplemental_Materials.docx`)S1–S3 KOSHA-aligned consolidated class-mapping rules (incl. KOSHA 24→8 accident-type mapping)S4 Baseline algorithm comparisonFig. S1 + S5 Classifier Chain auxiliary analysisS6 Multi-class cross-validation performanceS7 Progress-rate ablationS8 Binary-classifier SHAP Top-5 comparisonExtended methodological detail and the full consolidated class-mapping dictionaryReproducibility code (`Python_Code/`)Reproduces the leakage-safe feature engineering, KOSHA-aligned class consolidation, the multi-classdeterminant models, the three-tier SHAP attribution, the four-step reliability validation, theprogress-rate ablation, the binary high-risk classifier, the temporal-split validation, and theexternal validation against KOSHA. See `Python_Code/README.md` for environment and run instructions.---Data availabilityThe primary firm-level near-miss records (8,979 reports from 27 residential sites) are proprietary andconfidential under a non-disclosure agreement and are NOT included in this deposit. The code is designedto run on the original workbook when access is granted under the NDA.The KOSHA national construction-sector microdata (2017–2023) used for external validation are publiclyavailable at:https://portal.kosha.or.kr/archive/indus-acc-statis/indus-status-dataThis deposit contains only the supplemental document, the consolidated class-mapping dictionaries, and theimplementation code — none of which compromise data confidentiality.---How to cite> Lee, J. (2026). *Supplemental materials and reproducibility code for "From Reactive Records to Proactive> Indicators: Explainable Near-Miss Diagnostics in Construction."* Zenodo. https://doi.org/10.5281/zenodo.XXXXXXXX(Replace the DOI above with the one issued by the repository after publishing this record.)---LicenseCode (`Python_Code/`): MIT License (see `Python_Code/LICENSE`).Supplemental document and mapping dictionaries: Creative Commons Attribution 4.0 International (CC BY 4.0).