Journal of Chemometrics最新文献

Amomum tsao-ko Crevost et Lemaire (A. tsao-ko) is an important medicinal plant and flavoring spice. A. tsao-ko dried at different drying temperatures has different nutritional and medicinal values, leading to the phenomenon of substandard products in the market from time to time. In this study, attenuated total reflection–Fourier transform infrared spectroscopy (ATR-FTIR) data were pre-processed with SD, normalization, EWMA, SNV to compare their effects on the recognition ability of SVM, RF, XGBoost, and CatBoost models. Meanwhile, full-band and local-band 2DCOS profiles were obtained to characterize the differences in chemical features of A. tsao-ko dried by different drying temperatures and classified in conjunction with the ResNet model. The results show that although traditional machine learning can obtain better classification results, the classification efficiency is very unsatisfactory, and the correct classification rate is improved to 97% after derivative (SD) preprocessing. The 2DCOS atlas is able to visualize the feature information in the samples, which is further combined with the ResNet model to obtain 100% classification correctness with excellent generalization ability and convergence effect. The above study was able to provide new ideas for quality evaluation of A. tsao-ko.

The paper discusses methods of using chemometrics methods for processing the output data of sensors with polycomposite coatings for analyzing the gas phase of raw milk and obtaining analytical information about its total microbiological contamination, the content of yeast and mold, and the presence of pathogenic microorganisms. To predict microbiological indicators of milk quality, the partial least squares regression and quadratic discriminant analysis were used. The initial data matrix included both an optimized set of sensor output data and calculated parameters at various data fusion levels. It is shown that multidimensional patterns of sensor output data differ depending on the task. A model for predicting the microbiological contamination of milk (QMAFAnM) with an error of 0.342 log CFU was obtained. It was shown that the sensitivity of classification of milk samples by the presence or absence of pathogenic microorganisms using discriminant analysis is 67%, and the specificity is 100% when using the calculated parameters of the sensor array. The proposed approaches can be applicable for processing data from various types of sensors when analyzing real objects with complex compositions.

The discussions in the QSAR community and the steps that led to the definition of the OECD Principles for the validation of QSAR models are illustrated here, framing the process in the general framework of QSAR modeling. The individual OECD Principles are presented, commenting on them in the light of significant publications that have appeared over the years, with particular attention to the aspects of statistical validation according to the chemometric approach. It will be highlighted how and to what extent the OECD Principles have influenced the subsequent work of all QSAR modelers and have led to a significant improvement in validated QSAR modeling applicable in the regulatory field and beyond.

With the development of analytical instrument towards more and more high-way and complex, it is very important and meaningful work to obtain ultra-high-way chemical data and explore its analytical methods. In this paper, a novel and excellent six-way algorithm combination method (six-way ACM) was proposed. In addition, a real chemically meaningful ultra-high-way sexalinear data array was obtained and constructed for the first time. The proposed six-way data array has highly collinearity, which puts forward higher requirements for parsing this data array to a certain extent. To verify the feasibility of the proposed algorithm, it was used to analyze the above real sexalinear six-way data array and a series of simulated six-way data arrays with different noise levels. The results of real data and simulated data demonstrate that the proposed method can be well used in the analysis of six-way data arrays and shows fascinating performance, including insensitive to excessive number of components, fast convergence speed, and suitable for high collinearity and high noise data. Compared with three-way, four-way, and five-way calibration methods, the six-way ACM provides higher sensitivity, a lower limit of detection, a lower limit of quantification, and more stable and accurate results, showing an outstanding “higher-order advantages” and better ability to handle collinearity problems. This work provides not only data analysis method for high-order instruments that may emerge in the future but also real data support and methodological reference for theoretical research on high-order tensor algebra.

Tea color is a part of tea quality, and illegal addition of lead chrome green (LCG) to improve tea quality cannot be identified by human eyes. This paper is based on near-infrared (NIR) reflectance spectroscopy to detect LCG stained tea and to investigate the feasibility of qualitative and quantitative methods. Firstly, the LCG in tea was qualitatively analyzed by partial least squares discriminant analysis (PLS-DA), random forest (RF), and least squares support vector machine (LSSVM) classification models, and the results showed that the classification accuracy of LSSVM reached 100%. For quantitative analysis, Savitzky–Golay convolutional smoothing (S-G) preprocessing combined with three feature extraction algorithms, namely, joint competitive adaptive weighted sampling (CARS), uninformative variable elimination (UVE), and successive projection algorithm (SPA), were used to build partial least squares (PLS), RF, and LSSVM regression models sequentially on the preprocessed data. The S-G-UVE-LSSVM showed the best regression prediction ability in detecting LCG in tea, with a tested R² of 0.96. These results show the feasibility of NIR spectroscopy for the detection of added LCG in tea.

Halitosis is a condition associated with bad breath. Although halitosis is a disease in its own right, it is often a symptom of more serious diseases (diabetes mellitus, renal failure, azotemia, etc.). The currently used method for diagnosing halitosis is the organoleptic method, which relies on a trained specialist evaluating the patient's breath odor. This approach to diagnosing halitosis is subjective, uncomfortable for both patient and doctor, and necessitates the involvement of a specially trained professional. As an alternative, instrumental diagnostics employing metal oxide semiconductor (MOS) sensor arrays offer a promising avenue by enabling patient classification through predeveloped models. This paper considers the application of seven MOS sensors of different compositions at three different temperatures. Different methods of chemometric data analysis were applied: k-nearest neighbors (kNN), decision trees (DT), support vector machine (SVM), logistic regression (LR), and projection on latent structures discrimination analysis (PLSDA). All applied methods demonstrated their effectiveness and achieved selectivity, sensitivity, and accuracy values exceeding 85%. Additionally, a combined classifier leveraging responses from all previously studied classifiers was explored, achieving near-perfect classification accuracy.

In this paper, an innovative method for the simultaneous determination of nitrite, nitrate, and COD in water in the presence of turbidity as a source of noise in spectroscopic data has been investigated. UV–Vis absorption spectrometry and advanced machine learning are proposed to develop a stacking model, a sophisticated modeling approach that combines several basic models (PLS, Lasso, and Ridge regression) and a meta-regressor (Random Forest regressor) to improve prediction accuracy by incorporating baseline correction and principal component analysis (PCA) to mitigate the effects of turbidity on spectroscopic data. After applying these corrections, a significant improvement was observed: The root mean square error (RMSE) and the mean absolute error (MAE) were significantly reduced, and the correlation coefficient (R²) between predicted and actual values of nitrite, nitrate, COD, and turbidity was greater than 0.96, for all compounds in the test data set, that demonstrate the ability of the proposed stacking model to accurately predict nutrient concentrations simultaneously, even in complex environments; the proposed model may provide a valuable alternative to wet chemical methods. Due to its high accuracy and fast response, the proposed model can be used as an algorithm for the construction of nutrient sensors. This paper highlights the importance of integrating advanced modeling and data correction techniques to improve the robustness and accuracy of predictive models in environmental chemistry, thus providing valuable information for environmental monitoring and management.