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

Foodborne diseases (FBDs) pose a significant global public health challenge, necessitating rapid and accurate pathogen identification. However, real-world pathogen detection faces practical constraints, including few-shot scenarios for rare pathogens, severe class imbalance between common and rare classes, and the continuous emergence of novel pathogens requiring incremental learning capabilities. To address these challenges, this study proposes a Class Incremental Learning with Dual Attention and Adaptive Feature Fusion Network (CIL-DAFFNet). The model incorporates a dual-attention mechanism to enhance discriminative feature extraction from clinically imbalanced data and employs a dynamically weighted knowledge distillation strategy, adjusted via maximum mean discrepancy (MMD), to mitigate catastrophic forgetting. Evaluated on a real-world clinical dataset covering Salmonella, Vibrio parahaemolyticus, Escherichia coli, and Norovirus, CIL-DAFFNet outperformed five incremental learning baselines. In the final incremental phase, it achieved an accuracy of 0.8245, a Macro-F1 of 0.8004, and a G-mean of 0.7972. Ablation studies confirmed the contributions of core components, and SHAP analysis provided clinically interpretable feature importance rankings. This study provides a feasible technical solution for the rapid and precise identification of foodborne pathogens, supporting intelligent food safety monitoring and clinical decision-making.