An integrated machine learning framework for TCM five-flavor classification based on E-tongue profiling and SHAP analysis.

Abstract

Background: The "five-flavor" (wu wei) classification is a core organizing principle in Traditional Chinese Medicine (TCM) pharmacology and quality control, yet its assessment still relies on subjective organoleptic evaluation, limiting standardization and reproducibility.

Methods: Eighty-four processed herbal slices were analyzed using an electronic tongue to obtain multichannel sensor response profiles. Based on these data, we developed a dual-strategy ensemble learning framework in which a heterogeneous stacking model was trained for five-flavor identification while a voting ensemble was established in parallel for intensity grading. Model interpretability was evaluated using SHapley Additive exPlanations (SHAP) to quantify feature contributions and to relate prediction outcomes to specific electrochemical response patterns.

Results: On the independent test set, the stacking ensemble showed moderate and flavor-dependent classification performance, with a macro-AUC of 0.876 and a balanced accuracy of 0.629. Discriminative performance was higher for Bitter and Sweet than for Sour, Salty, and Pungent. Full-stack SHAP analysis further identified P_13 and P_8 as the most influential features overall, with P_13 contributing mainly to the discrimination of bitter and pungent attributes, whereas P_8 was more strongly associated with sour, salty, and sweet characterization.

Conclusions: This study demonstrates a data-driven workflow for digitizing TCM five-flavor assessment. Further validation with expanded sample sizes and chemical corroboration of sensor attributions is needed before operational deployment.