A Machine Learning Model to Reframe the Concept of Shelf-Life in Bakery Products: PDO Sourdough as a Technological...

Traditional shelf-life (SL) determination in bakery products relies primarily on subjective sensory evaluation, limiting both predictive capability and technological transfer. This study aimed to develop an objective, data-driven framework by integrating statistical and Machine Learning (ML) methods to identify and quantify the core determinants of bread SL. Samples were produced under a 2 × 2 × 2 factorial design (Fermentation, Temperature, Packaging), with continuous monitoring of physicochemical and atmospheric parameters. Three-way ANOVA confirmed that Storage x Temperature (η² ÷ 0.41) and Modified Atmosphere Packaging (η² ÷ 0.36) were the dominant factors. The optimal synergy (4 °C + ATM) achieved a 100% Success Rate, extending SL to 54 days vs. 16 days under ambient conditions. For prediction, a Generalized Linear Model (GLM) was developed for binary classification and rigorously validated via 10-fold cross-validation. The GLM achieved an Overall Accuracy of 89% (AUC 92%), uniquely identifying pH and Total Titratable Acidity (TTA) as the most influential predictors. In conclusion, GLM provides a robust tool for objective SL prediction. The integrated ANOVA-GLM framework achieved a 3.3-fold SL extension and 92% predictive accuracy. The findings confirm that preservative effectiveness is not solely due to the process itself, but is mediated by the resulting chemical acidity, offering a scalable framework for Real-Time Quality Control (QC) in the food industry.