Comprehensive analysis of tea concentrates with machine learning assisted sensory characterization

• Broken Orange Pekoe emerged as the best overall tea leaves for tea concentrate. • TPC, DPPH, and Tannins were the strongest predictors of sensory attributes. • SVR Outperformed XGB and RF in Predicting Sensory Quality. This study investigated physicochemical, biochemical, microbiological, and sensory changes in seven tea concentrate types during 30-day frozen storage (-18°C), and developed machine learning models to predict sensory quality from chemical parameters. Tea concentrates prepared from FBOP, FP, BOP, GBOP, CD, RD, and PF leaves underwent progressive transformations: total phenolic content declined 19–30%, theaflavins decreased 70–80%, while highly polymerized substances increased 7–15%. Sample 3 (Broken Orange Pekoe) achieved the highest sensory scores (38.76/45, "Excellent") through balanced polyphenol retention, while Sample 7 exhibited the poorest performance (35.31/45, "Moderate") due to severe degradation. All samples remained microbiologically safe (bacterial/fungal counts <10⁵ CFU/mL, no E. coli). Support Vector Regression demonstrated robust predictive capacity (R²=86.83–98.73%) for estimating sensory attributes from chemical composition. SHAP analysis revealed that total phenolic content, DPPH, and tannins consistently dominated sensory predictions, while theaflavins, despite traditional emphasis, showed relatively lower importance in concentrated matrices. This suggests polymerized phenolic products compensate for theaflavin loss during storage. The findings indicate that rapid chemical assays (TPC, DPPH, tannins) could serve as proxies for sensory assessment, potentially reducing reliance on resource-intensive sensory panels for quality monitoring.