Non-destructive detection of mercury and salicylic acid in beauty cream using hyperspectral imaging and chemometrics for toxicological risk assessment

Abstract

Adulteration of cosmetic products with toxicants remains a public-health concern, especially in markets where surveillance is uneven. We developed a non-destructive screening method for mercury and salicylic acid using hyperspectral imaging (400–1000 nm, Specim FX10 camera) coupled with machine-learning (ML) chemometrics. A total of 1,800 hyperspectral image samples (covering nine classes, including pure and graded adulteration) were collected. Since each hyperspectral image produces a large volume of spectral information, preprocessing was applied to extract usable data. This process resulted in approximately 45,000 images spanning 224 spectral bands/features, which were then used for machine learning–based predictive analysis. The spectra were rigorously preprocessed using the empirical line method for radiometric calibration and Savitzky–Golay smoothing before comparative modelling with histogram-based gradient boosting (H-GB), artificial neural networks (ANN), one-dimensional convolutional neural networks (1D-CNN), and one-vs-one linear discriminant analysis (OvO-LDA). To the best of our knowledge, this is the first application of HSI integrated with a 1D-CNN for cosmetic adulterant screening, highlighting its novelty within this domain. The 1D-CNN delivered the best overall performance, with validation accuracy up to 97% (10-fold CV mean 0.940 ± 0.015), while OvO-LDA was the most stable across folds; H-GB showed lower performance for low-level mercury adulteration. Spectral analysis highlighted 500–750 nm as the most discriminative region for both analytes, informing future band selection. The approach enables rapid, real-time screening without destroying samples, supporting regulatory spot-checks and factory-floor QC.