Machine learning-assisted detection of Sudan I in chili powder using FT-IR coupled with magnetic molecular imprinting extraction

This study presents a rapid and selective analytical method for the quantification of Sudan I in chili powder by integrating magnetic molecularly imprinted solid-phase extraction (MISPE) with machine learning (ML)-assisted Fourier transform infrared (FTIR) spectroscopy. The key methodological innovation lies in a ratio-based feature engineering strategy, where peak areas of Sudan I characteristic bands (1600, 1550, and 1350 cm -1 ) are normalized against Fe 3 O 4 reference bands (590 and 425 cm -1 ) originating from the magnetic core. This approach effectively compensates for variability in sample mass and KBr pellet preparation—a long-standing but previously unaddressed challenge in quantitative FTIR analysis. Notably, the method requires no elution step; Fe 3 O 4 @SiO 2 @MIP S nanoparticles with adsorbed Sudan I are directly mixed with KBr for FTIR measurement, where Fe 3 O 4 simultaneously serves as the adsorbent carrier and an internal standard. A systematic comparison of ten regression algorithms identified Random Forest as the optimal model, achieving R 2 = 0.938, RMSE = 7.459 mg/L, and RPD = 4.017 on an independent test set. Method validation using spiked chili powder samples yielded satisfactory recoveries (92.0–97.0%) with good precision (RSD < 9.5%), and the complete workflow from sample extraction through ML-based prediction was completed within 90 minutes. The proposed method bridges the traditional trade-off between the accuracy of chromatographic confirmatory techniques and the speed of rapid screening tools, offering a practical platform for high-throughput, lower-cost monitoring of illicit dyes in food commodities, with potential adaptation to portable FTIR instrumentation for on-site applications.