PredictingNext-Inspection Cooling-Tower Regulatory Noncompliance to Support Inspection Prioritization: Evidence from New York City

Cooling-tower noncompliance can weaken controls intended to limit Legionella-related public-health risks. We examined whether longitudinal inspection histories could predict at least one violation at a system’s next inspection and support capacity-constrained prioritization. We reconstructed 109,913 retained inspection-result records from the New York City Department of Health and Mental Hygiene, covering January 2017 through March 2026, into 35,825 inspections of 6162 systems. After excluding each system’s first inspection, we modeled 29,663 repeat inspections of 5668 systems using 19 history-only predictors. We compared logistic regression, XGBoost, and TabM under system-grouped random and chronological designs, assessed discrimination and probabilistic accuracy, examined model reliance using permutation importance, and retrospectively evaluated prioritization at 10%, 20%, and 30% capacity. The grouped-random test set contained 4217 inspections, of which 1295 were noncompliant; average precision ranged from 0.6695 to 0.6852 and AUROC from 0.7864 to 0.8003. At 10% capacity, the models captured 26.56–27.26% of noncompliant inspections and 42.33–42.64% of violations, with lifts of 2.65–2.72. Performance declined on 5421 chronologically held-out inspections (890 noncompliant): average precision ranged from 0.2645 to 0.2939 and AUROC from 0.6221 to 0.6490. At 10% capacity, and the models captured 18.76–22.02% of noncompliant inspections and 24.05–26.23% of violations, with lifts of 1.87–2.20. The models generally only modestly outperformed four simple history-based rankings. Permutation importance indicated strongest reliance on prior violation burden, inspection-history depth, and elapsed-time measures. Historical records supported retrospective prioritization, but performance weakened for later inspections. Prospective evaluation is needed before operational use; predictions concern recorded regulatory noncompliance, not Legionella contamination or direct public-health risk.