Heterogeneous ensemble learning applied to UV-VIS identification of multi-class pesticides by high-performance liquid chromatography with diode array detector (HPLC/DAD)

Abstract

While infrared and mass spectral libraries are well documented, the same cannot be said for the ultraviolet and visible region (UV-VIS), severely impacting HPLC/DAD identification operations. Considering advancements in machine learning and its technologies, the exhaustive task of compiling and maintaining extensive standardized libraries may become obsolete. When well-tuned to the problem of spectral recognition, machine learning models can identify complex patterns and relationships within spectra, reducing the need for direct comparison with reference spectra. Therefore, this study proposed the development and validation of a heterogeneous ensemble model, integrating decision tree algorithms and meta-learning techniques, specialized in UV-VIS spectral recognition using HPLC/DAD. The ensemble demonstrated satisfactory performance, with an accuracy of 95.88 ± 4.45 % and a precision of 96.74 % (MCC = 0.9571) with data from the test set (n = 97), and an accuracy of approximately 80 %, but with a considerable recall of 93.00 %, when evaluated with real application data. A weighted quantitative index based on the feature importance parameter of random forests was developed and applied to estimate the model probability of success. The model, its constituents and other additional resources were made available in an open repository.