Interpretable Machine Learning for Predicting Enrofloxacin Residues in Fish Using a Large Literature-Derived Database

Enrofloxacin use in aquaculture can lead to residue accumulation in fish tissues, creating challenges for food safety control and residue risk assessment. Residue prediction across published studies remains difficult because concentrations are influenced by species, tissue matrix, administration conditions, environmental factors, and between-study heterogeneity. Here, we constructed a literature-derived database to model matrix-specific enrofloxacin concentrations in fish. After duplicate removal, fish-focused filtering, source cleaning, and exclusion of records with missing critical variables, 1254 records from 39 source groups with traceable literature identifiers were retained from 2275 extracted records. Eleven predictors were used, and eight machine learning models were trained using log1p-transformed concentrations. Under a random row-level 80:20 split, the histogram gradient-boosting decision tree achieved the best row-level performance, with R2_log = 0.8889 and root-mean-squared error in log1p space (RMSE_log) = 0.2728. In contrast, stratified source-grouped validation showed markedly reduced performance, indicating limited cross-study generalization to unseen source groups. ExtraTrees showed the most stable grouped validation performance and was used for model-based interpretation. Species, tissue or biological matrix, dosing frequency, and sampling time point were highly ranked model-associated predictors. This study provides a transparent literature integration framework for residue pattern modeling and predictor prioritization, while highlighting the need for standardized residue depletion data and external validation.