Rapid Assessment of Quality Changes in French Fries during Deep-frying Based on FTIR Spectroscopy Combined with Artificial Neural Network

: Fourier transform infrared (FTIR) spectroscopy combined with backpropagation artificial neural network (BP-ANN) were utilized for rapid and simultaneous assessment of the lipid oxidation indices in French fries. The conventional indexes (i.e. total polar compounds, oxidized triacylglycerol polymerized products, oxidized triacylglycerol monomers, triacylglycerol hydrolysis products, and acid value), and FTIR absorbance intensity in French fries were determined during the deep-frying process, and the results showed the French fries had better quality in palm oil, followed by sunflower oil, rapeseed oil and soybean oil. The FTIR spectra of oil extracted from French fries were correlated to the reference oxidation indexes determined by AOCS standard methods. The results of BP-ANN prediction showed that the model based on FTIR fitted well (R 2 > 0.926, RMSEC < 0.481) compared with partial least-squares model (R 2 > 0.876). This facile strategy with excellent performance has great potential for rapid characterization quality of French fries during frying.

and the main fractions of polar compounds. However, these conventional methods are laborious and time-consuming with chemicals and fail to achieve the facile application in the frying industry 9,10 .
For rapid and simultaneous determination of lipid oxidation in food, Fourier transform infrared FTIR spectroscopy can provide a solution for this challenge. FTIR spectroscopy was widely applied in the food detection due to environmental-friendliness, low-cost, and tiny sample and easy preparation for determination 11 13 . FTIR has also been used to determine TPC content of frying oil 14 16 . Artificial neural network ANN is an algorithm that models non-linear relationships by mimicking the functionality of a neural network, which contains 3 parts of the input layer, the hidden layer and the output layer and is typically specified by architecture, learning algorithm and neuron model. ANN has many advantages, e.g able to handle data with high collinearity and noise, able to model complex nonlinear relationships between independent and dependent variables, Training algorithm allows neural network to self-learn , able to detect interactions between predictor variables, requires less formal statistical training to develop 17 . ANN has been applied successfully to the automation and intelligent control of food drying process 18 , and food authentication 19,20 , due to its self-learning ability, adaptive nature, strong fault tolerance 21,22 . Karaman et al. estimated the oxidation parameters peroxide value, free fatty acids and iodine value of sunflower oil added with some natural byproduct extracts combined with ANN 23 . However, to the best of our knowledge, there is limited research about FTIR spectroscopy combined with artificial neural network for the rapid and simultaneous determination of oxidation indexes in French fries.
In this study, FTIR spectroscopy in association with the back propagation artificial neural network BP-ANN model was developed for the routine analysis of TPC, TGP, THP, OTG, and AV of French fries fried in palm oil F-PO , sunflower oil F-SuO , soybean oil F-SO and rapeseed oil F-RO . The feasibility and sensitivity of the method were evaluated and compared with the partial least-squares PLS model. The FTIR spectroscopy method will provide an alternative approach for monitoring the quality changes of French fries rapidly, contributing to the safety of fried food and scientific frying during deep-frying.

Materials and reagents
Fresh potatoes Solanum tuberosum L, Helan15 cultivar , were obtained from a farm in Shandong Province, China. The frying experiments were conducted with soybean oil SO , rapeseed oil RO , palm oil PO , and sunflower oil SuO without additives donated by Wilmar, and stored at 4 for further use. Acetone, diethyl ether, petroleum ether, toluene, hexane analytical reagent grade and Tetrahydrofuran chromatography grade were purchased from Sinopharm Chemical Regent Co., Ltd Shanghai, China and Sigma St. Louis, MO, USA , respectively.

Preparation of French fries and oil extraction from
French fries A stainless deep fryer Aigoli, China with a maximum capacity of 2.5 L was used for frying. Potatoes were washed and cut into pieces 1 1 6 0.5 cm with a manual cutter. Firstly, 2.0 L of fresh oil was heated to 168 3 in the fryer. Then, a batch of 100 2 g of potato strips was fried in oils regularly. Frying time 4 min and waiting time 26 min were set in each frying batch. The overall time was 3 days 8 h/day . Frying oils SO, RO, PO, and SuO at different frying time of 0.06, 2,4,6,8,10,12,14,16,18,20,22, and 24 h were used for French fries frying. Extraction of oil from the French fries was performed according to the method described by Li et al., 8 . Those samples were stored at 20 until analysis.

The physicochemical properties of French fries
The fatty acid composition was analyzed by a gas chromatography Agilent 7820, USA Table S1 . The conditions were performed according to our previous research 24 . AV of samples were measured according to the AOCS official method 3a-63 25 . The determination of TPC was based on the AOCS official method Cd 20-91 26 . The compositions of TPC e.g., TGP, OTG, THP in oils extracted from French fries were obtained by a high-performance size-exclusion chromatography method described by our previous study 27 .

FTIR analyses
A Bruker FTIR spectrometer VERTEX 70 series was employed to obtain the FTIR spectra. Each recorded spectrum was obtained by averaging 16 scans at a resolution of 4 cm 1 . Each sample was measured three times. The average of the three spectra obtained from each same sample was used in subsequent analysis. The spectra file of samples was pretreated through normalization to reduce measurement error. Each spectrum was aligned to the baseline using instrumentation software. 100 μL of oil extracted from the French fries was deposited onto PE film surface to form a uniform film. With PE film as background, the spectrum of oil was obtained. Using the calibration equation associating the effective path length with the absorbance of 4334 cm 1 , the effective path lengths of spectra were normalized to 0.15 mm 28 .

Artificial neural network modeling and partial leastsquares model
The BP-ANN is a multi-layer feedforward neural network trained based on the error back propagation algorithm 29 . It is composed of an input layer, hidden layer, and output layer. In general, the three-layer network can approach the target value with arbitrary precision if the number of hidden layers is sufficient. It is typically specified by architecture, learning algorithm and neuron model, in which architecture represents the interconnection pattern between the different layers of neurons, learning algorithm is for updating the weights in order to correctly model a particular task and neuron model defined by activation function is to transform a neuron s weighted input to its output activation. Through parameter optimization in the pre-experiment, the best BP-ANN model was obtained with the neuron number of 10 in a hidden layer, the training function of trainlm, and the transfer functions of tansig and purelin in hidden layer and output layer, respectively. Therefore, a three-layer network was selected for all neural network models in this study, and 156 sets of data were obtained and divided into training data 70 , 110 samples , and validation data 15 , 23 samples , and testing data 15 , 23 samples . Based on our preliminary experiment and previous study 30 , the FTIR absorbance peak at 1696 cm 1 , 3471 cm 1 and 968 cm 1 related to oil oxidation were used as inputs of BP-ANN model to predict the conventional indexes output TPC, TGP, THP, OTG, and AV . The partial least-squares PLS model was also validated using the FTIR absorbance peak height at 1696 cm 1 , 3471 cm 1 and 968 cm 1 and these relevant conventional indexes. Root Mean square error RMSE , coefficient of determination R 2 of the predicted conventional indexes were computed to evaluate the performance of fitting and predicting. Furthermore, in order to compare the quality of BP-ANN model and PLS, root mean square error of calibration RMSEC and root mean square error of prediction RMREP were calculated for the evaluation. Besides 10 blind French fries samples were collected as external validation samples.

Statistical analyses
The reference values and the spectra collected of the samples were conducted in triplicate. The relevant results were denoted as mean standard deviation SD . BP-ANN was performed using MATLAB R2014a The MathWorks Inc., Natick, MA, USA . The PLS model was performed using SIMCA-P version 13.0, Umea, Sweden . Spectral data analyses were conducted with the use of OMNIC 7.3 Thermo Electron Inc., Madison, WI .

Results of chemical analysis
Before FTIR analysis, the TPC, TGP, THP, OTG, and AV of oil samples from French fries were determined. Here, the degree of thermal oxidative decomposition in French fries fried in frying oils at 0.06, 4,8,12,16,20 h are displayed in Fig. 1.
With a longer period of the frying process, the TPC in French fries kept increasing but the increase rates declined gradually in general. For TPC, the content of polar compounds in F-SO was consistently higher than that in F-RO, F-SuO, and F-PO, the polar compounds in F-SuO and F-PO was the lowest during frying. The F-SO reached 26.64 at 20 h frying compared with F-PO with TPC of 19.61 .  While, the OTG contents greatly fluctuated in F-SuO and F-PO and their contents were lower than F-SO and F-RO at the primary frying period before 15 h Fig. 1C . The THP contents in French fries increased exponentially with the frying time. The order of THP average growth rates were F-SO 0. 22 F-RO 0.17 F-SuO 0.15 F-PO 0.12 Fig. 1D . The AV increased with the frying time as presented in Fig. 1E, and F-SO presented a higher value compared to other samples.

FTIR analysis
FTIR spectroscopy is a rapid method for food detection, which is an alternative to traditional methods in the food industries. The complete FTIR spectra of oil samples extracted from French fries were exhibited in Fig. 2A. For lipid oxidation analysis, the most essential spectral regions correspond to ROOH, FFA C O bond , and -HC CHtrans functional groups 31 . More specifically, the corresponding regions were the following wavenumbers: the -HC CH-trans bond occurred at nearly 980-960 cm 1 ; the FFA C O bond occurred between 1720-1680 cm 1 and the ROOH corresponded to the regions of 3650-3300 cm 1 31 33 . The change in intensity of these spectral regions of oil samples could reflect the chemicals transformation during frying, consequently denoting the degree of lipid oxidation.
Functional group spectra s variation for samples in the course of frying was shown in Fig. 2A. The FFA C O bond absorbance at about 1696 cm 1 , -HC CH-trans bond absorbance at about 968 cm 1 and ROOH characteristic absorption peak height at about 3471 cm 1 of oils extracted from French fries showed the growth trend during frying. The starting adjacent spectral gaps at the height of these absorption peaks were small, whereas the adjacent spectral gaps increased with time, suggesting that hydrogen peroxide, free fatty acids and isolated trans fatty acids increased with time Figs. S1 A-C . Figures 2B-2D showed that F-PO and F-SuO displayed the lower characteristic absorption peak height at 3471, 1696 and 968 cm 1 compared with F-SO and F-RO by comparing the FTIR ab- Fig. 2 Change of the functional group spectra for French fries sample fried in SO A , changes of absorbance peak at 3471cm 1 B , 1696 cm 1 C , 968 cm 1 D for French fries fried in different oils during frying.
sorbance intensity of the ROOH, FFA C O bond , and -HC CH-trans bond in French fries samples, the results revealed that the degree of oxidative stability was F-PO F-SuO F-RO F-SO.

Model establishment and veri cation by BP-ANN
To establish the BP-ANN model for the change of quality in French fries during deep-frying, the FTIR absorbance peak at 1696, 3471 and 968 cm 1 , were set as variable X. Meanwhile, the conventional indexes TPC, TGP, THP, OTG, AV were set as variables Y. Through parameter optimization, the BP-ANN model with the topology of 3-10-1 the neuron number of the input layer, hidden layer, and output layer was 3, 10, and 1, respectively , performing algorithm of Levenberg-Marquardt, the transfer function of tansig and purelin in hidden layer and output layer, respec-tively, and training function of trainlm Levenberg-Marquardt outperformed the fitting performance and accuracy. Training set included 70 of the samples, and the prediction models of TPC, TGP, THP, OTG, and AV established by BP-ANN method are shown in Fig. 3. The predicted values were well correlated with the actual values, and the R 2 of the training set were all 0.915. 15 of the samples were used as the validation set and 15 as a test set. The results showed that the R 2 of the validation set and testing set of TPC, TGP, THP, AV were 0.945 and 0.940, respectively, indicating that the prediction ability of the model was available Table 1 , numerous studies revealed that the R 2 higher than 0.9 in the model indicated excellent classification performance as well as prediction ability 34 . However, R 2 of the testing set of OTG was only 0.863, which is not good owing to the fluctuation of OTG   3 -68.9 , and the R 2 value was increased by 4.14 -10.15 , respectively. In our study, the lower RMSEC, RMSEP and higher R 2 of the calibration training equations for TPC, TGP, THP, OTG, AV were obtained compared with PLS models. The BP-ANN had achieved excellent predictive performance R 2 0.926 for the TPC, TGP, THP, AV conventional indexes during the frying process.

Sample external validation
After calibrating the models, external validation procedure was carried out to provide an estimate of the overall accuracy of the predictions. Blind samples from 10 different French Fries samples prepared in different oxidation degree of oils were used to validate the performance of the proposed method. The TPC, TGP, THP, OTG, and AV of the oils were initially analyzed using AOCS standard method before the proposed FTIR analysis. The results obtained are presented in Fig. 4. Figure 4 confirms that the two methods agreed with each other well, with R 2 0.975 for TPC, TGP, THP, and AV prediction, except for OTG R 2 0.656 . FTIR methods combined with ANN could measure efficiently the TPC, TGP, THP, and AV of various French fries. Therefore, the proposed method can be used as an alternative TPC, TGP, THP, and AV determination method.

Conclusions
This study aimed to simplify oxidation parameters determination in French fries using a rapid and convenient PE film-based FTIR procedure, which are considered practical for routine implementation. Furthermore, compared with PLS model, the BP-ANN model showed good performance and excellent prediction for the changes of quality TPC, TGP, THP, and AV in French fries during frying based on FTIR spectroscopy. FTIR spectral regions correspond to ROOH, FFA C O bond , and -HC CH-trans functional groups were selected for the model establishment. With the increase of the FTIR absorbance intensity in these regions, a deeper degree of thermal oxidative decomposition was presented in French fries during frying. The proposed method indicated that the quality of F-PO was better than F-SuO and F-RO, followed by F-SO. This investigation opens new opportunities for developing FTIR spectroscopy combined with BP-ANN as a promising method for the qualitative detection of fried food in a wide range of applications.

Acknowledgement
This work was financially supported by the National First-Class Discipline Program of Food Science and Technology JUFSTR20180202 ; Postgraduate Research & Practice Innovation Program of Jiangsu Province KYCX20_1852 ; National Natural Science Foundation of China 31901728 ; Jiangsu Planned Projects for Postdoctoral Research Funds 2020Z297 . We thank Jianhua Huang, Ting Shi, and Zhe Dong for assisting in preparation of this manuscript.