Machine learning-assisted multimodal lateral flow immunoassay based on urchin-like Au@Pt nanoparticles for quantitative determination of febuxostat in functional foods.

BACKGROUND: The illegal addition of febuxostat (FEB) to functional foods poses a significant food safety risk and calls for rapid and reliable analytical methods. Conventional single-signal lateral flow immunoassays (LFIAs) often show limited quantitative reliability in complex matrices because each signal mode is affected by matrix interference and modality-dependent bias. RESULTS: A machine learning-assisted multimodal LFIA based on urchin-like Au@Pt nanoparticles was developed, which simultaneously generated colorimetric (catalytic), fluorescence, and photothermal readouts on a single strip. The rough Au@Pt surface improved antibody coupling efficiency from 76.0% for AuNPs to 94.1%. The three complementary signals were integrated using a supervised multilayer perceptron regression model for FEB quantification. Under repeated random train/test split evaluation, the multimodal model achieved a mean test R2 of 0.72 ± 0.09 and a mean test MAE of 4.51 ± 0.89 μg/kg, and outperformed unimodal and baseline regression models. Among the individual modes, the assay achieved limits of detection as low as 0.31 μg/kg (fluorescence quenching mode) and a working range up to 0.34-40.0 μg/kg (photothermal mode). Combined with the MLP regression model, the multimodal assay provided recoveries of 99.7%-101.9% and CVs below 3% in spiked samples. SIGNIFICANCE: This study demonstrates that integrating complementary multimodal LFIA signals with supervised regression can reduce modality-dependent error and improve quantitative reliability. The proposed strategy provides an effective framework for quantitative FEB screening in complex supplement matrices and supports the development of more reliable machine learning-assisted immunoassays for food safety analysis.