Multimodal machine learning reveals strain-specific flavor and kinetic dynamics in chinese spicy cabbage fermentation.

This study presents a novel multimodal machine learning framework integrating physicochemical kinetics with volatile metabolomics to elucidate strain-specific fermentation dynamics in Chinese spicy cabbage (CSC). Over a 90-day fermentation under ten single-strain LAB inoculation conditions, the modified Weibull model accurately described distinct growth-decline patterns, categorizing strains into three physiological types: slow-growing, robust-survival, and fast-growing. By integrating multimodal data including organic acids, reducing sugars, and volatile organic compounds (VOCs), multiple machine learning models coupled with SHAP analysis identified key discriminative indicators across fermentation stages, revealing a pronounced time-dependent shift in feature importance. In the early fermentation stage, strain differences were primarily distinguished by VOCs, particularly volatile sulfur compounds, reflecting strain-specific amino acid metabolism. In contrast, during the late storage stage, physicochemical indicators such as lactic acid concentration, LAB counts, total titratable acidity, and pH effectively differentiated strains, highlighting varied long-term acid production and survival capacities. Furthermore, partial least squares analysis and kernel-based change point detection revealed that CSC fermentation exhibited the fastest rate of change within the first 1-3 days, with a kinetic transition detected at 5-7 days coinciding with the protocol-driven temperature shift from 10 °C to 4 °C, marking the transition from a rapid dynamic phase to a relatively stable phase under the employed industrial fermentation conditions. This integrated framework provides a comprehensive temporal and strain-specific reference for rational starter culture selection and targeted process regulation in long-term fermented vegetable production.