Developing robust food composition models: Strategies for handling temperature and packaging variations in dry-cured...

Low-cost near infrared devices intended for consumers able to easily determine composition and quality of food products may boost adoption of sustainable healthy diets. However, predictive algorithms robust to external variations are needed. The aim of this work was to evaluate different data analysis strategies to develop robust predictive models for food composition when using spectrometric data subjected to external variations, specifically temperature and packaging material, acquired using low-cost sensors. Usefulness of global modelling (GM), Generalised least squares weighting (GLSW), Loading space standardisation (LSS), Multiplicative Effects Model (MEM) were explored, and the effect of samples heterogeneity evaluated. To do so, two low-cost handheld NIR-based devices with different spectral ranges and resolutions were used. The food matrix samples were obtained from different anatomical muscles of commercial dry-cured ham. Spectra were acquired on two types of packaging films at different temperatures to further explore the usefulness of global modelling (GM), generalised least squares weighting (GLSW), loading space standardisation (LSS), and multiplicative effects model (MEM) to retrieve these effects. Results show that the inherent food sample heterogeneity produces as much spectral variability as temperature and packaging materials. For temperature compensation, LSS did not decrease the predictive error caused by this factor probably due to the heterogeneity of the samples used. In contrast, the GLSW method decreased the predictive errors from 0.52% to 0.46% for salt and from 2.10% to 1.40% for water.. Only a slight effect of packaging was observed, and GM models were found to be the best strategy to compensate it, showing a decrease of bias from -1.35 to 0.012. The examined compensation strategies could facilitate the deployment of low-cost spectrometers for consumer use, as they offer an effective means to mitigate or eliminate variations from any source in the data that are unrelated to the properties of interest.