Rapid and non-destructive detection of formaldehyde adulteration in shrimp based on deep learning-assisted portable Raman spectroscopy

Abstract

Formaldehyde (FA), a known carcinogen, is occasionally used illegally as a preservative in seafood, while traditional detection methods for FA residues often fail to meet the practical needs for nondestructive detection. In this study, a approach was developed by combining a portable Raman spectrometer with the InceptionTime deep learning model without sample pretreatment. Model were trained by FA-negative and FA-positive Raman spectral data from the shrimp surface and achieved accuracies of 84.40 % and 85.17 % at detection thresholds of 5 mg/kg (the primary safety detection threshold) and 100 mg/kg (the abuse-level contamination threshold), respectively. Metabolomic analysis and weight visualization indicated that the model particularly focused on Raman peaks associated with specific amino acids and astaxanthin-binding proteins. Two amino acid metabolites, timonacic and spinacine, were also identified as direct indicators of FA addition. Our model offers a field-deployable and practical approach for real-time and on-site FA detection scenario.