Multivariate integration of physicochemical and targeted metabolite profiles with sensory data for the predictive modeling of taste intensities in fermented kimchi.

Conventional kimchi quality control relies on trained sensory panels, limiting throughput and hindering the establishment of chemically grounded criteria linking composition to taste across fermentation stages. A chemically interpretable framework was developed to estimate five sensory attributes (sourness, saltiness, sweetness, umami, and spiciness) by integrating panel scores with routine physicochemical indices and targeted metabolite quantification. Thirty-two commercially available kimchi products were evaluated at three predefined refrigerated-storage time points, generating 96 stage-wise observations spanning early, optimal, and over-ripened fermentation stages defined by pH criteria. This design captured both between-product heterogeneity and within-product fermentation progression. Sourness exhibited the strongest stage dependency, whereas sweetness and umami followed non-monotonic trajectories and spiciness remained comparatively stable. Linear discriminant analysis based on sensory profiles and compositional measurements consistently separated fermentation stages, indicating stage-dependent chemical signatures across products. False discovery rate-adjusted correlation analysis identified taste-linked markers, with titratable acidity strongly associated with sourness (r = 0.934) and capsaicin strongly associated with spiciness (r = 0.807). Taste-specific regression equations combining physicochemical variables with selected targeted markers showed the strongest apparent performance for sourness (R2 = 0.948), followed by spiciness (R2 = 0.790) and saltiness (R2 = 0.709), whereas sweetness and umami showed more moderate apparent performance (R2 = 0.438-0.471). Prediction bias remained small across attributes (|Bias| ≤ 0.139). Predictive performance was further evaluated using repeated product-level grouped cross-validation, in which all stage-wise observations derived from the same commercial product were assigned to the same fold. In addition, the established equations were applied without coefficient re-estimation to 28 additional commercially distributed kimchi samples. This applicability assessment retained meaningful observed-predicted agreement for saltiness, spiciness, and sourness, whereas sweetness showed limited external agreement. These findings demonstrate that sensory attributes with relatively well-defined chemical drivers can be estimated from measurable compositional markers within the sampled commercial kimchi domain. The present study provides a chemically interpretable basis for composition-based sensory taste estimation and fermentation-stage quality assessment in commercially distributed kimchi.