Machine learning-driven discovery of key aroma discriminants in stale- and betelnut-aroma Liupao tea: Integration of E-nose and GC-IMS

A machine learning-driven strategy for optimal feature selection is proposed. • W1W and W1C sensors serve as key variables for aroma type classification. • 131 volatile compounds were identified in Liupao tea. • 14 aroma type differentiating compounds were identified. • 4 key variables yielded 100% accurate aroma recognition models. This study aimed to differentiate between the stale-aroma and betelnut-aroma types of Liupao tea by analyzing their volatile profiles. An integrated methodology was employed, combining electronic nose (E-nose) technology with headspace gas chromatography-ion mobility spectrometry (HS-GC-IMS). The E-nose data effectively differentiated between the two aroma types in both dry leaves and tea infusions. HS-GC-IMS enabled identification of 131 volatile compounds. Key discriminative features were selected through machine learning by integrating the variable importance in projection (VIP) from orthogonal partial least squares discriminant analysis (OPLS-DA) with Shapley additive explanations (SHAP) values from extreme gradient boosting (XGBoost). This process identified two critical E-nose sensor channels and two key volatile compounds. Utilizing these four optimal variables, three classification models were developed: principal component analysis-linear discriminant analysis (PCA-LDA), partial least squares discriminant analysis (PLS-DA), and support vector machine (SVM). All models demonstrated a perfect classification accuracy of 100% on both the training and internal test datasets. The PCA-LDA and PLS-DA models maintained 100% accuracy on the independent test set, whereas the SVM model achieved a classification accuracy of 80% on the same set. These results reveal the volatile characteristics of Liupao tea and offer valuable insights for quality assessment and targeted flavor modulation.