Naslov
Artificial neural network models for determination of histamine in fish by Surface Enhanced Raman Spectroscopy

Autori
Janči, Tibor ; Valinger, Davor ; Gajdoš Kljusurić, Jasenka ; Mikac, Lara ; Ivanda, Mile ; Vidaček, Sanja

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Proceedings of 47th WEFTA conference / - , 2017, 125-125

Skup
47th Conference of the West European Fish Technologists’ Association (WEFTA) conference

Mjesto i datum
Dublin, Irska, 09.10.2017. - 12.10.2017

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
histamine ; fish ; SERS ; artificial neural network

Sažetak
In our previous work Surface Enhanced Raman Spectroscopy (SERS) method for determination of histamine has been developed. Although SERS offers a possibility of rapid analysis with minimal sample preparation, interpretation of recorded spectra often isn't straightforward and requires spectra pre-processing by different mathematical algorithms to eliminate side effects such as fluorescence background, detector noise and cosmic spikes. Optimization of parameters for spectral pre- treatment can be time consuming and if not conducted properly can have negative effect on predictive quality of chemometric models leading to incorrect results. In this work, application of artificial neural network (ANN) as a mathematical tool for analysis of large sets of data, i.e. unprocessed SERS spectra, and prediction of histamine content in range 0 – 400 mg/kg was examined. Nonlinear neural multiple layer perceptron (MLP) network was applied for prediction of histamine content and 2 ANN models were developed. ANN-1 model was developed on the basis of 10 individual SERS spectra for each concentration of histamine while ANN-2 model was developed on the basis of average of 10 recorded spectra for each histamine concentration. One hidden layer was chosen for ANN development and the number of neurons in the hidden layer was set to a range from 3 to 11. Selection of the optimal neural network architecture was performed by comparing the values of the root mean square error (RMSE) and the linear correlation coefficient (R2). In terms of correlation coefficient, better results were obtained for model ANN-2 (R2 = 0, 978) compared to ANN-1 (R2 = 0, 906). Model ANN-2 performed significantly better and in wider concentration range (0 – 400 mg) compared to previously developed partial least square regression model (R2 = 0, 962, concentration range 0 – 200 mg/kg). Obtained results confirm that developed model ANN-2 can be successfully used for determination of histamine content in fish samples by SERS in concentration range 0 – 400 mg/kg.

Izvorni jezik
Engleski

Znanstvena područja
Prehrambena tehnologija

POVEZANOST RADA