Temporal Screening of High-Risk Food Service Inspections in New York State, 2023-2025: A Case Study Using Multimodal...

Food safety inspection systems generate rich historical records, yet converting these records into actionable pre-inspection risk signals remains challenging under limited regulatory resources. The objective of this study was to develop and evaluate a temporally valid, leakage-free, multimodal screening framework for identifying high-risk food service inspections before the upcoming inspection outcome is known. Existing studies have improved inspection prediction with machine learning, but many focus on contemporaneous classification rather than temporally valid high-risk screening, and few jointly model historical numeric behavior, prior narrative context, and predictive uncertainty. To address this gap, this study proposes a temporal high-risk food inspection screening framework based on multimodal evidential learning. Using New York State food service inspection data, we constructed a event-level dataset of 55,454 inspections from 20,082 establishments and predicted whether an upcoming inspection would be high-risk using only pre-inspection information. The proposed evidential deep learning multilayer perceptron integrates current metadata, longitudinal numeric history, and historical inspection comments while producing calibrated uncertainty estimates for selective prediction. On the held-out test set, the proposed model achieved the best overall performance, with an AUROC of 0.846, AUPRC of 0.424, F1 score of 0.431, Brier score of 0.063, and ECE of 0.012, outperforming strong tabular baselines including CatBoost and TabM. Under selective prediction, its retained-set F1 increased from 0.431 at full coverage to 0.542 at 80% coverage. Explainability analysis further showed that predictive gains were driven primarily by historical compliance dynamics, with historical text providing complementary contextual value. These findings support the use of temporally valid, uncertainty-aware multimodal models for risk-based food inspection prioritization.