Benchmarking machine learning methods for comprehensive chemical fingerprinting and pattern recognition

• Published in: Journal of Chromatography A Vol. 1595; pp. 158 - 167
• Main Authors: Reichenbach, Stephen E., Zini, Claudia A., Nicolli, Karine P., Welke, Juliane E., Cordero, Chiara, Tao, Qingping
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 21.06.2019
• Subjects: Benchmarking; Chemistry Techniques, Analytical - methods; Classification; Comprehensive two-dimensional gas chromatography; Data mining; GCxGC; Machine Learning; Pattern Recognition, Automated - methods; GCxGC; Data mining; Comprehensive two-dimensional gas chromatography; Machine learning; Classification
• ISSN: 0021-9673; 1873-3778
• DOI: 10.1016/j.chroma.2019.02.027

•Relatively simple machine learning (ML) methods can be highly accurate with GCxGC.•Chemical fingerprints from smart templates capture information-rich patterns.•Untargeted TIC profiles can be nearly as useful as target quantifier ions. Machine learning (ML) has been used previously to recognize particular patterns of constituent compounds. Here, ML is used with comprehensive chemical fingerprints that capture the distribution of all constituent compounds to flexibly perform various pattern recognition tasks. Such pattern recognition requires a sequence of chemical analysis, data analysis, and pattern analysis. Chemical analysis with comprehensive multidimensional chromatography is a maturing approach for highly effective separations of complex samples and so provides a solid foundation for undertaking comprehensive chemical fingerprinting. Data analysis with smart templates employs marker peaks and chemical logic for chromatographic alignment and peak-regions to delineate chromatographic windows in which analytes are quantified and matched consistently across chromatograms to create chemical profiles that serve as complete fingerprints. Pattern analysis uses ML techniques with the resulting fingerprints to recognize sample characteristics, e.g., for classification. Our experiments evaluated the effectiveness of seventeen different ML techniques for various classification problems with chemical fingerprints from a rich data set from 126 wine samples of different varieties, geographic regions, vintages, and wineries. Results of these experiments showed an accuracy range from 58% to 88% for different ML methods on the most difficult classification problems and 96% to 100% for different ML methods on the least difficult classification problems. Averaged over 14 classification problems, accuracy for the different methods ranged from 80% to 90%, with some relatively simple ML techniques among the top-performing methods.