Mike Hardy, Hossein Kashani Zadeh, Angelis Tzouchas, Fartash Vasefi, Nicholas MacKinnon, Gregory Bearman, Yaroslav Sokolov, Simon A Haughey, Christopher T Elliott
Salmon fillet was analyzed via hand-held optical devices: fluorescence (@340 nm) and absorption spectroscopy across the visible and near-infrared (NIR) range (400-1900 nm). Spectroscopic measurements were benchmarked with nucleotide assays and potentiometry in an exploratory set of experiments over 11 days, with changes to spectral profiles noted. A second enlarged spectroscopic data set, over a 17 day period, was then acquired, and fillet freshness was classified ��1 day via four machine learning (ML) algorithms: linear discriminant analysis, Gaussian na��ve, weighted -nearest neighbors, and an ensemble bagged tree method. Dual-mode data fusion returned almost perfect accuracies (mean = 99.5 �� 0.51%), while single-mode ML analyses (fluorescence, visible absorbance, and NIR absorbance) returned lower mean accuracies at greater spread (77.1 �� 10.1%). Single-mode fluorescence accuracy was especially poor; however, via principal component analysis, we found that a truncated fluorescence data set of four variables (wavelengths) could predict "fresh" and "spoilt" salmon fillet based on a subtle peak redshift as the fillet aged, albeit marginally short of statistical significance (95% confidence ellipse). Thus, whether by feature selection of one spectral data set, or the combination of multiple data sets through different modes, this study lays the foundation for better determination of fish freshness within the context of rapid spectroscopic analyses.
