Assessment of Sauvignon Blanc Aroma and Quality Gradings Based on Static Headspace-Gas Chromatography-Ion Mobility Spectrometry (SHS-GC-IMS): Merging Analytical Chemistry with Machine Learning

Abstract

<div>In this paper, we report on the application of the static headspace-gas chromatography-ion mobility spectrometry (SHS-GC-IMS) instrument in the field of wine aroma analysis and its potential in constructing a prediction model for the quality gradings of wines. The easy-to-operate, cost effective SHS-GC-IMS instrument was innovatively used for a non-targeted search for volatile compounds in Sauvignon Blanc wine, with the identification of volatiles seldom before reported. The wine aroma profile acquired by the instrument was organically and innovatively combined with advanced classification models, inspired by the computer science community, to produce high classification accuracy in terms of wine quality gradings. Useful insights were also extracted by using advanced interpretation methods on complex models to learn the important volatiles correlated with wine quality grading.</div>

Non-Targeted Screening Approaches for Profiling of Volatile Organic Compounds Based on Gas Chromatography-Ion Mobility Spectroscopy (GC-IMS) and Machine Learning

“…Other supervised methods used in NTS with HS-GC-IMS are gradient boosting (e.g., XGBoost) [31], decision tree classification (Tree) [91], logistic regression (Regressor) [91], orthogonal partial least-squares discriminant analysis (OPLS-DA), quadratic discriminant analysis (QDA) [30], or soft independent modeling of class analogy (SIMCA) [82]. Furthermore, nonlinear classifications are often performed using support vector machines (SVMs).…”

“…To prevent overfitting, the data set is split into training and test data. The ratio between training and test data, which is commonly referred to as 'train-test split', is usually between 2:1 [88] and 4:1 [33], and sometimes as low as 6:1 [31]. The test set is used to determine the prediction accuracy, which is usually lower than the classification accuracy but more meaningful [88].…”

“…Leaveone-out CV leaves out a single observation at a time, while k-fold CV splits the data into k subsets, which are one by one held out as the validation set [107]. For NTS using HS-GC-IMS, 10-fold CV is commonly performed [31,52], alongside with leave-one-out CV [80,85]. Bootstrapping is a resampling method that can be used as an alternative to CV to estimate the prediction performance of a model with a low number of training samples.…”

“…ANN has recently the gained attention for applications in food science and technology, while often being limited by the requirement of having sufficiently large data sets [116,117]. Zhu and coworkers have shown the superior prediction performance of ANN (89.5%) over PCA-LDA (65.7%), PLS-DA (58.7%), kNN (k = 5, 60.8%) and SVM (51.8%), and XGBoost (81.8%), for the classification of Sauvignon Blanc via SHS-GC-IMS [31]. Vega-Márquez and coworkers evaluated a deep learning network and five different benchmark methods for the classification of olive oil samples into EVOO, OO, and LVOO, based on HS-GC-IMS spectra obtained for 701 olive oil sample from two different harvests [91].…”

“…According to the European Union directive, the exemption limit for the total activity for tritium was set to 1 GBq [29]. Therefore, the usage of a low-radiation tritium ion source with an activity of 300 MBq is not subject to authorization, hence leading to a broad adoption of tritium ion sources in a number of commercially available systems on the market [30][31][32][33][34]. Beta particles, which are emitted by the tritium source, initiate a gas-phase reaction cascade of the drift gas (nitrogen or air), resulting in predominant proton-water clusters H + [H 2 O] n , which are commonly referred to as 'reactant ions' [35].…”