Machine Learning-Assisted Multicolor Fluorescence Assay for Visual Data Acquisition and Intelligent Inspection of Multiple Food Hazards Regardless of Matrix Interference.

Abstract

Regarding the significant health risks of pesticide residue in foods, while current sensors still suffer from limited efficiency and stability, as well as difficulties in qualitative identification and quantitative detection of mixtures, development of innovative detection techniques combined with advanced methodology holds great research value. Herein, a highly efficient intelligent food risk evaluation system was proposed by integrating a multicolor fluorescent responsive assay with machine learning (ML) algorithms for the identification and quantification of multiple pesticides, carbendazim (CBZ), heptachlor (HEP), chlordimeform (CDF), and their mixtures. This method leveraged the color changes generated from the interaction between multicolor carbon dots (CDs) and target pesticide molecules. By extracting color signal feature values from these reactions and integrating the visual data acquisition with ML models, this method enables efficient qualitative identification and quantitative detection of multiple pesticides, regardless of matrix interference through a dual-source data acquisition strategy without large instruments. The developed evaluation system via a ″stepwise prediction″ strategy automatically demonstrated robust qualitative identification capability with a discrimination accuracy of 99.3% for pesticide categorization while achieving robust quantitative prediction accuracy (R2 ≥ 0.8946) for pesticide concentration detection, verified in six kinds of food matrix. This method significantly improves the detection stability and efficiency, providing a promising tool for food safety monitoring.