Escrever uma avaliação PREreview

Building code waiver assessments in Singapore remain largely discretionary, relying on case officers' subjective judgment with limited decision-support tooling. This study presents the first machine learning framework for predicting building code waiver outcomes, trained on 197 historically decided cases from the Building and Construction Authority (BCA) across five waiver categories: barrier-free accessibility (n = 45), ventilation (n = 61), staircase design (n = 37), safety provisions (n = 30), and structural modifications (n = 24), spanning 2021 to 2023. Fourteen engineered features, including documentation completeness, technical justification quality, and compliance history, were extracted through domain-expert annotation. Four models were evaluated: L2-regularised logistic regression, random forest, gradient boosting (XGBoost), and a weighted ensemble. The ensemble achieved the highest predictive accuracy of 83.7% (95% CI: 79.2-88.1%) with an area under the receiver operating characteristic curve (AUC) of 0.891 (95% CI: 0.854-0.928), significantly outperforming all individual models (McNemar's test, p < 0.05). SHAP analysis revealed that documentation completeness and technical justification quality collectively account for 55% of prediction variance. A companion five-by-five risk assessment matrix, combining predicted rejection probability with consequence severity, stratified cases into actionable risk tiers correlating with observed approval rates from 90.3% (very low risk) to 10.0% (very high risk; Spearman rho = -0.71, p < 0.001). The framework offers a transparent, data-driven complement to regulatory judgment and demonstrates feasibility for integration into Singapore's Corenet X digital building submission platform.