Effects of High hydrostatic pressure Treatment on the Flavor Quality of Kiwi Wine Based on Machine Learning and Flavoromics

To investigate the effect of high hydrostatic pressure (HHP) treatment on the volatile flavor profile of kiwi wine, this study applied HHP at 300, 400, and 500 MPa for 10 and 20 min, with an untreated control, yielding seven treatment groups. Each group comprised three independent samples (n = 3) for machine learning analysis, with five and three technical replicates for GC-MS and E-nose, respectively. A flavoromics approach integrating gas chromatography-mass spectrometry (GC-MS), electronic nose (E-nose), and machine learning was employed; volatile compounds were identified by matching mass spectra against the NIST library, and data were analyzed by one-way ANOVA, principal component analysis (PCA), and multi-class classification models. To prevent data leakage, all preprocessing steps were performed independently within each fold of stratified 3-fold cross-validation repeated 100 times, with one sample per group assigned to the test set and the remaining two to the training set per fold. A total of 40 volatile compounds were identified by GC-MS, among which esters, alcohols, and ketones collectively accounted for over 90% of the total relative content, constituting the principal volatile components of kiwi wine. HHP treatment significantly modulated volatile composition in a pressure- and time-dependent manner: the 300 MPa-20 min treatment increased phenethyl acetate and ethyl laurate while decreasing ethyl decanoate, suggesting a potential shift toward floral and fruity aroma notes based on volatile compound profiles; the 400 MPa-10 min treatment induced a transient accumulation of acetaldehyde that decreased with extended treatment time; whereas the 500 MPa-20 min treatment reduced ester contents and accumulated higher alcohols such as isoamyl alcohol, causing a pronounced deviation from the original volatile profile. E-nose PCA effectively distinguished the control from treatment groups with different pressure-time combinations. Among logistic regression, linear SVM, RBF-SVM, and LDA models, the RBF-SVM model exhibited the best classification performance, achieving accuracies of 0.905 and 0.762 for GC-MS and E-nose data, respectively. These findings contribute to the understanding of HHP-induced flavor changes in kiwi wine and offer preliminary reference data for the optimization of non-thermal processing parameters in fruit wine production.