Spectroscopic methods and machine learning modelling to differentiate table eggs from quails fed with different inclusion levels of silkworm meal

The study aimed at evaluating the VIS/NIR and NIR instrument's capability to discriminate among table eggs from quails fed with different inclusion levels of silkworm (Bombyx mori L.) pupa meal (SWM). The trial consisted of four experimental groups of laying quails fed for 8 weeks with a 0%, 4%, 8% or 12% SWM inclusion levels. At week 7, 120 eggs (30 per experimental group) were sampled to form the training set used to perform the spectroscopy-based classification models, whereas at the end of the trial a second batch (n = 48) was used as an independent test set to assess the reliability of the classification models. Using a benchtop and two portable devices, VIS-NIR and NIR spectra were recorded from the training set and subsequently submitted to a random forest (RF) feature selection. The selected NIR informative wavelengths were used to perform the following supervised classification algorithms: partial least squares-discriminant analysis (PLS-DA), K-nearest neighbour (KNN) and support vector machine (SVM) linear and radial. The RF features predictive models were applied against the independent test set to assess the reliability by a set of confusion matrices. A moderate predictive capacity of the tested VIS/NIR and NIR devices in predicting the eggs feeding groups was observed probably due to the relative low difference in SWM amount among the classes. After a merging of the SWM4-12 experimental groups, both NIR benchtop and portable devices combined with the advanced machine learning models successfully showed capacity to recognize in the eggs the inclusion of insect meal in layers diet since an accuracy higher than 0.90 has been observed. Instead, the VIS-NIR portable tool recorded worse predictive capacity with an accuracy lower than 0.73. The most informative NIR wavelengths belonged to the 1350–1600 and 1850–2200 nm spectral regions. The achieved outcomes in terms of accuracy suggested that both KNN and SVM models provided more powerful machine learning algorithm than PLS-DA. The results showed that a portable NIR spectrometer had comparatively accurate classification to the benchtop instrument, highlighting the potential of hand-held NIR spectrometers in at-line monitoring eggs from SWM-fed layer quails.