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

Area of Science:

• Food Safety
• Machine Learning
• Public Health

Background:

• Food safety inspection systems generate extensive historical data.
• Translating this data into effective pre-inspection risk signals is difficult due to limited regulatory resources.
• Existing machine learning models often lack temporal validity and fail to integrate diverse data types for risk screening.

Purpose of the Study:

• To develop and assess a temporally valid, multimodal screening framework for identifying high-risk food service inspections.
• To create a leakage-free system that predicts inspection outcomes using only pre-inspection data.
• To improve upon existing methods by jointly modeling historical numeric data, narrative context, and predictive uncertainty.

Main Methods:

• A temporal high-risk food inspection screening framework based on multimodal evidential learning was proposed.
• An evidential deep learning multilayer perceptron was used, integrating metadata, longitudinal numeric history, and inspection comments.
• The model was trained and tested on a dataset of 55,454 New York State food service inspections.

Main Results:

• The proposed model achieved superior performance with an AUROC of 0.846 and AUPRC of 0.424.
• It outperformed strong tabular baselines like CatBoost and TabM.
• Selective prediction enhanced the F1 score from 0.431 to 0.542 at 80% coverage, demonstrating effective risk prioritization.

Conclusions:

• The developed framework provides a temporally valid, uncertainty-aware approach for risk-based food inspection prioritization.
• Multimodal evidential learning effectively integrates diverse data sources for improved prediction accuracy.
• The findings support the use of advanced AI models to optimize regulatory resource allocation in food safety.