Class-Incremental Learning for Foodborne Pathogen Prediction Based on Clinical Surveillance Data.

Area of Science:

• Microbiology
• Computer Science
• Public Health

Background:

• Foodborne diseases (FBDs) present a major global health concern, requiring swift and precise identification of causative agents.
• Current pathogen detection methods struggle with challenges like limited data for rare pathogens, imbalanced datasets, and the need to adapt to new pathogen discoveries.

Purpose of the Study:

• To develop an advanced Class Incremental Learning (CIL) model, the CIL-DAFFNet, designed to overcome limitations in real-world foodborne pathogen identification.
• To enhance the accuracy and adaptability of pathogen detection systems for improved food safety and clinical diagnostics.

Main Methods:

• Proposed the Class Incremental Learning with Dual Attention and Adaptive Feature Fusion Network (CIL-DAFFNet).
• Integrated a dual-attention mechanism for improved feature extraction from imbalanced clinical data.
• Utilized a dynamically weighted knowledge distillation strategy with Maximum Mean Discrepancy (MMD) to prevent catastrophic forgetting in incremental learning.

Main Results:

• CIL-DAFFNet demonstrated superior performance against five incremental learning baselines on a real-world clinical dataset.
• Achieved high accuracy (0.8245), Macro-F1 score (0.8004), and G-mean (0.7972) in the final incremental phase.
• Ablation studies and SHAP analysis validated the model's effectiveness and provided interpretable feature importance.

Conclusions:

• CIL-DAFFNet offers a robust solution for the rapid and accurate identification of foodborne pathogens, addressing key real-world challenges.
• The model supports intelligent food safety monitoring and aids clinical decision-making by providing reliable pathogen detection.
• This research advances the field of incremental learning for public health applications, particularly in combating foodborne illnesses.