Smart screening of acrylamide in biscuits using NIR spectroscopy and machine learning modeling: a PAT-oriented study

Acrylamide (AA) is a process contaminant formed during high-temperature treatment of carbohydrate-rich foods, raising major toxicological concerns and prompting stringent regulatory limits. This study investigated the potential of optical sensing techniques as rapid and non-destructive tools for AA screening in cocoa-based biscuits. Three spectroscopic devices (VNIR, NIR, and FT-NIR) were evaluated across two industrial sampling campaigns, comprising 98 ground biscuit samples collected directly from the production line immediately after baking in a continuous oven. To maximize process variability, seven production factors were systematically modified. Biscuit AA concentrations, determined by LC-MS/MS, ranged from 100 to 500 & micro;g/kg. The developed PLS regression models achieved cross-validation R2 between 0.65 and 0.87 and residual prediction deviation (RPD) values up to 2.0, supporting their use for semi-quantitative screening. Domain shifts between campaigns introduced spectral variability primarily associated with differences in storage conditions. Application of external parameter orthogonalization (EPO) effectively compensated for such variability, improving external prediction performance (RMSE of prediction approximate to 40-50 & micro;g/kg). These results demonstrate that VNIR/FT-NIR spectroscopy can serve as a Process Analytical Technology tool for at-line or in-line monitoring of AA in bakery production, supporting a Quality-by-Design framework through faster process feedback and proactive control of food safety and quality.