engleski

Assessment of microbiological status of chilled salmon fillets using Raman spectroscopy and machine learning

Microbial activity is the main mechanism of quality loss during refrigerated storage of fish, affecting a wide array of parameters such as sensory and functional characteristics of fish muscle, development of undesirable smell and formation of biogenic amines, due to fish muscle degradation. Traditional microbiological colony count methods are often laborious and time consuming, requiring long incubation time to obtain results, making them inconvenient for use in high-throughput quality control laboratories. In this work, application of Raman spectroscopy as a rapid, noninvasive method for determination of aerobic mesophilic count (AMC) of salmon fillets was examined. For this purpose, 25 salmon fillets were stored in refrigerator at temperature of 4°C and periodically analysed by Raman spectroscopy and traditional colony counting method. Five Raman spectra of each fillet were recorded using process – type Raman Rxn2 Hybrid analyzer equipped with PhAT optical probe (Kaiser Optical Systems Inc.) and 785 nm excitation wavelength laser in spectral range from 150–1875 cm-1. Obtained spectra were pre-processed and scaled and fed to machine learning models by means of partial least square (PLS) regression, Random Forests (RF) regression and deep neural networks (DNN). The predictive features (X matrix) were the processed Raman spectra per sample, while the target variable (y) was AMC of the sample. Best results were obtained using PLS regression model with cross validation parameters Rcv2=0.85 and root mean square error RMSECV=0.85. Besides model training, this study also utilizes the paradigm of post-hoc model explanation based on permutation importance. Most important features in this study are based on 1305-1318, 1651-1658 and 995-1005 cm- 1 spectral regions.

Raman spectroscopy ; machine learning ; microorganisms ; fish ; quality

nije evidentirano