Explainable Transformer-Based Modelling for Pathogen-Oriented Food Safety Inspection Grade Prediction Using New York...

Area of Science:

• Public Health
• Computer Science
• Food Safety

Background:

• Foodborne pathogens pose significant public health risks.
• Early identification of unsafe food conditions is crucial for prevention.
• Routine inspections generate valuable data for risk assessment.

Purpose of the Study:

• To develop an explainable transformer-based framework for predicting food safety inspection grades.
• To utilize multimodal inspection data, combining structured metadata and unstructured deficiency narratives.
• To evaluate the performance of various machine learning and deep learning models, including transformers.

Main Methods:

• Combined structured metadata with unstructured deficiency narratives.
• Evaluated classical machine learning (LightGBM), deep learning (BiLSTM), and transformer models (RoBERTa).
• Employed SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• RoBERTa achieved the highest performance with an F1 score of 0.96.
• BiLSTM and LightGBM also showed strong performance (F1=0.95 and F1=0.92, respectively).
• SHAP analysis identified key indicators of pathogen-related hazards, including temperature abuse, pests, and unsanitary practices.

Conclusions:

• Transformer-based models with explainable AI (XAI) can effectively support pathogen-oriented monitoring and real-time risk assessment.
• Multimodal AI approaches can enhance inspection efficiency and strengthen public health surveillance.
• The developed framework shows potential for improving food safety and reducing foodborne illnesses.