Machine learning-based sensor array: full and reduced fluorescence data for versatile analyte detection based on gold nanocluster as a single probe

Different acquisition data approaches have been used to fetch the fluorescence spectra. However, the comparison between them is rare. Also, the extendability of a sensor array, which can work with heavy metal ions and other types of analytes, is scarce. In this study, we used first- and second-order fluorescent data generated by 6-Aza-2-thiothymine-gold nanocluster (ATT-AuNCs) as a single probe along with machine learning to distinguish between a group of heavy metal ions. Moreover, the dimensionality reduction was carried out for the different acquisition data approaches. In our case, the accuracy of different machine learning algorithms using first-order data outperforms the second-order data before and after the dimensionality reduction. For proving the extendibility of this approach, four anions were used as an example. As expected, the same finding has been found. Furthermore, random forest (RF) showed more stable and accurate results than other models. Also, linear discriminant analysis (LDA) gave acceptable accuracy in the analysis of the high-dimensionality data. Accordingly, using LDA in high-dimensionality data (the first- and second-order data) analysis was highlighted for discrimination between the selected heavy metal ions in different concentrations and in different molar ratios, as well as in real samples. Also, the same method was applied for the anion's discrimination, and LDA gave an excellent separation ability. Moreover, LDA was able to differentiate between all the selected analytes with excellent separation ability. Additionally, the quantitative detection was considered using a wide concentration range of Cd2+, and the LOD was 60.40 nM. Therefore, we believe that our approach opens new avenues for linking analytical chemistry, especially sensor array chemistry, with machine learning.

Automated summary

This study used fluorescence spectra data and machine learning to distinguish heavy metal ions, and found that the accuracy of first-order data is higher than second-order data. Linear discriminant analysis performed well in high-dimensional data analysis and could differentiate ions with different concentrations and molar ratios.