Proactive early invisible spoilage-risk prediction through uncertainty-quantified temporal microbial growth intelligence

The study addresses the need for proactive food safety monitoring because microbial growth can increase before visible spoilage becomes apparent. Existing microbial prediction methods mainly provide deterministic forecasts and often lack temporal interpretability, predictive uncertainty, and decision-oriented risk estimation. This study proposes the Bayesian Attention Temporal Microbial Growth Network (BAT-MG Net) to forecast microbial concentration and estimate early spoilage risk. The model is implemented using Python with GPU-accelerated deep-learning tools, using ComBase microbial growth time-series data. Missing values are handled using training-derived imputation, features are Min-Max normalized, and group-stratified 80:20 partitioning prevents experimental leakage. A 6-h temporal window and a 6-h-ahead forecasting horizon are used with LSTM and attention mechanisms, while Monte Carlo Dropout quantifies predictive uncertainty. The resulting D-PSRI combines microbial concentration and uncertainty to classify risk as low, moderate, or high. BAT-MG Net achieved 97.8% accuracy, 0.198 RMSE, 0.154 MAE, and 0.97 R 2 , with 94.6% PICP and 0.021 calibration error. These results indicate effective microbial concentration forecasting and uncertainty-aware spoilage-risk estimation within the evaluated ComBase analytical subset. Previously reported studies are included as contextual references because differences in datasets, prediction endpoints, experimental conditions, and modeling tasks prevent direct head-to-head comparisons of performance.