Physics‑informed machine learning modeling with dual SHAP analysis for predicting Salmonella Enteritidis growth on chicken meat treated with slightly acidic electrolyzed water

Abstract

Slightly acidic electrolyzed water (SAEW) is an effective sanitizer for controlling Salmonella Enteritidis in chicken processing, but existing mechanistic models suffer from systematic prediction bias due to structural simplifications, limiting its precise application under complex temperature scenarios. To address this bias, we developed a SGompertz–CatBoost residual-correction hybrid model and conducted dual SHAP analyses to diagnose bias sources and interpret decision mechanisms. The hybrid model achieved an R 2 of 0.981 and reduced RMSE by 41.2% relative to the mechanistic model. Residual SHAP attributed the systematic bias to three structural deficiencies: the symmetric sigmoidal assumption, the linear temperature–available chlorine concentration (ACC) interaction, and the linear temperature response. Hybrid SHAP further verified that residual correction shifted the mechanistic prediction ( y phys ) from the primary bias source to a stable baseline; time–temperature interaction terms ( tT, tT 2 , and t 2 T ) autonomously captured nonlinear thermal kinetics; ACC of SAEW emerged as an independent antimicrobial signal; and the contributions of mechanistic parameters collapsed to near zero. Scenario-based prediction indicated that 30 mg/L ACC reduced predicted S. Enteritidis peaks by 1.5 – 1.9 log CFU/g across storage scenarios. Superiority probability heatmaps revealed that higher ACC SAEW consistently exhibited probabilistic dominance at 4 °C (0 – 350 h), 14 °C (0 – 130 h), and 24 °C (0 – 90 h). This study provides a robust framework for SAEW process optimization and risk assessment under controlled isothermal storage conditions.