Microbiological Quality Assessment of Chicken Thigh Fillets Using Spectroscopic Sensors and Multivariate Data Analysis

Fourier transform infrared spectroscopy (FT-IR) and multispectral imaging (MSI) were evaluated for the prediction of the microbiological quality of poultry meat via regression and classification models. Chicken thigh fillets (n = 402) were subjected to spoilage experiments at eight isothermal and two dynamic temperature profiles. Samples were analyzed microbiologically (total viable counts (TVCs) and Pseudomonas spp.), while simultaneously MSI and FT-IR spectra were acquired. The organoleptic quality of the samples was also evaluated by a sensory panel, establishing a TVC spoilage threshold at 6.99 log CFU/cm2. Partial least squares regression (PLS-R) models were employed in the assessment of TVCs and Pseudomonas spp. counts on chicken’s surface. Furthermore, classification models (linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machines (SVMs), and quadratic support vector machines (QSVMs)) were developed to discriminate the samples in two quality classes (fresh vs. spoiled). PLS-R models developed on MSI data predicted TVCs and Pseudomonas spp. counts satisfactorily, with root mean squared error (RMSE) values of 0.987 and 1.215 log CFU/cm2, respectively. SVM model coupled to MSI data exhibited the highest performance with an overall accuracy of 94.4%, while in the case of FT-IR, improved classification was obtained with the QDA model (overall accuracy 71.4%). These results confirm the efficacy of MSI and FT-IR as rapid methods to assess the quality in poultry products.


Introduction
Food waste amounts to 14% of the worlds' food consumption [1], while meat and specifically poultry production is forecasted to rise at 137 million tones [2]. In addition, consumer's awareness and demand for high quality and safety meat and poultry has been continuously increased. For this purpose, non-invasive spectroscopic sensors have been used in the evaluation of the quality and freshness of meat products [3] through the implementation of process analytical technology (PAT) [4,5]. The underlying principle of PAT is to combine spectral data acquired through real-time (in-, on-, at-line) non-destructive analytical techniques with multivariate data analysis for the development of models assessing food quality. These models, along with their datasets, could be uploaded in the cloud, updated regularly with new data in order to be consultative to the food industry [6].
In recent years, multispectral imaging (MSI) and Fourier transform infrared (FT-IR) spectroscopy have been investigated as alternative methods for the evaluation of a variety of meat products [7][8][9]. The former method is a merge of UV and NIR with computer vision, and it has been proposed as an ecological approach for rapid quality evaluation of meat and poultry [10][11][12]. Until now, spectral data in the visible and near-infrared region (400-1700 nm) have been employed in the development of quantitative or qualitative models for the determination of the bacterial population (TVCs and Pseudomonas spp.) on chicken meat during spoilage [13][14][15]. In the same context, MSI analysis has been proved a solution to the identification of adulteration/food fraud of minced beef with chicken meat [16], as well as for the detection of food fraud in minced pork adulterated with chicken [17]. Moreover, fecal contaminants in poultry line [18] and the presence of tumors on the surface of chicken breasts [19] have been accurately detected via MSI analysis. This innovative method was successfully employed in the at-line estimation of the time from slaughter in four different poultry products [20].
Likewise, the potential of FT-IR for the qualitative and quantitative assessment of the microbiological quality of meat products has been explored by other researchers [21][22][23][24][25]. Especially for poultry, FT-IR was recommended as an effective approach for the differentiation of intact chicken breast muscle during spoilage [22]. Additionally, the level of spoilage bacteria on the surface of chicken meat was successfully estimated via FT-IR spectroscopy [21]. Further investigation of this promising method for real-time evaluation of the freshness of stored chicken breast fillets was undertaken by Vansconcelos et al. [26]. FT-IR analysis was also proposed as an efficient approach for the categorization of chicken meat among seven raw types of food, irrespective of variations among batches and storage conditions (temperature, storage duration, packaging, spoilage levels) [3].
Spectral data acquired by nondestructive methods such as MSI and FT-IR have been analyzed by a variety of unsupervised and supervised machine learning algorithms for the rapid quality assessment in food matrices including meat [25,27,28]. Partial least squares regression (PLS-R), linear discriminant analysis (LDA), and quadratic discriminant analysis (QDA) have been reported as reliable tools for the development of predictive models for spoilage or adulteration assessment in meat [9,[29][30][31]. Moreover, deep learning methodologies such as artificial neural networks (ANNs) and support vector machines (SVMs) [32] have been employed, validated, and compared through available websites (e.g., sorfML, Metaboanalyst) or softwares (R, MatLab, Python), in an attempt to provide accurate quantitative and qualitative models for food spoilage assessment [28,[33][34][35][36][37].
The aim of the present work was to develop and evaluate machine learning regression (PLS-R) and classification models (LDA, QDA, SVMs, QSVMs) based on MSI and FT-IR spectral data for the evaluation of the microbiological quality of chicken thigh fillets. More specifically, PLS-R models were developed for the prediction of the microbiota of TVCs and Pseudomonas spp. on the surface of chicken thigh, whereas LDA, QDA, SVMs, and QSVMs models were employed for the classification of samples in two quality classes (fresh or spoiled) based on the outcome of sensory analysis. The challenging task in this study was not confined in model development, batch variation and different storage temperatures, but it also considered external validation using two different dynamic temperature profiles simulating temperature scenarios during transportation and storage in retail outlets.

Experimental Design
Three hundred and thirty (330) chicken thigh fillets (ca. 90-110 g/fillet) enclosed in plastic packages (dimensions = 25 cm (width), 90 µm (thickness), permeability ca. 25, 90, and 6 cm 3 m −2 day −1 bar −1 at 20 • C and 50% RH for CO 2 , O 2 , and N 2 , respectively) were obtained from a poultry industry in Greece and stored aerobically at eight isothermal conditions (0, 5,10,15,20,25,30, and 35 • C). Two independent experiments were undertaken at all isothermal conditions using 4 different batches of chicken meat. Moreover, 72 samples were stored at two dynamic temperature profiles (profile 1 = 12 h at 5 • C, 8 h at 10 • C, and 4 h at 15 • C; profile 2 = 12 h at 0 • C, 8 h at 5 • C, and 4 h at 10 • C), simulating temperature scenarios that can be observed during transportation and storage in retail outlets [38]. At pre-determined time intervals, packages were subjected to microbiological analyses for the enumeration of total viable counts (TVCs) and Pseudomonas spp., in parallel with MSI and FT-IR spectral data acquisition. At each sampling point, duplicate packages per isothermal storage condition and triplicate packages from each dynamic temperature profile were subjected to the abovementioned analyses. In addition, chicken samples were subjected to sensory evaluation by a 14-member sensory panel to categorize the samples in two quality classes, namely fresh and spoiled as detailed below. Microbiological counts and sensory scores were correlated with spectral data in order to develop quantitative and qualitative models assessing chicken thigh's microbial loads (TVCs, Pseudomonas spp.) as well as their quality class (fresh-spoiled).

Microbiological Analysis and Sensory Evaluation
A total surface of ca. 20 cm 2 (four slices of ca. 5 cm 2 each with a maximum thickness of 2 mm) from chicken thigh fillet was removed aseptically, by means of a sterile stainless steel cork borer (2.5 cm in diameter), scalpel and forceps, added in 100 mL of sterile quarter strength Ringer's solution (Lab M Limited, Lancashire, UK) and homogenized in a Stomacher device (Lab Blender 400, Seward Medical, UK) for 120 s at room temperature [39]. The microbial load on the surface of chicken was enumerated using serial decimal dilutions in the same Ringer's solution and 0.1 mL of the appropriate dilution was spread on the following growth media: (a) Tryptic glucose yeast agar (Plate Count Agar, Biolife, Milan, Italy) for the determination of total viable counts (TVCs) incubated at 25 • C for 72 h; (b) Pseudomonas agar base (LAB108 supplemented with selective supplement Cetrimide Fucidin Cephaloridine, Modified C.F.C. X108, LABM) for the determination of presumptive Pseudomonas spp. incubated at 25 • C for 48 h. The results were logarithmically transformed and expressed as log CFU/cm 2 .
In parallel, sensory evaluation was performed by a 14 member in-house trained sensory panel. For this purpose, samples (n = 103) were placed in sterile petri dishes and scored according to their odor using a 3-point hedonic scale as follows-1 = fresh, 2 = acceptable, 3 = spoiled [40]. Samples with scores < 2 were characterized as fresh (Class 1) whereas samples with scores ≥ 2 as spoiled (Class 2). Finally, the sensory outcome was correlated with spectral data in order to assess the quality class of the samples directly from the acquired MSI and FT-IR spectra.

Spectra Acquisition
Multi-spectral images (MSI) were captured using a Videometer-Lab instrument (Videometer A/S, Herlev, Denmark) that acquires images in 18 different non-uniformly distributed wavelengths from UV (405 nm) to short wave NIR (970 nm), namely, 405, 435, 450, 470, 505, 525, 570, 590, 630, 645, 660, 700, 850, 870, 890, 910, 940, and 970 nm. Detailed information about this spectroscopic sensor is provided elsewhere [41]. Each sample corresponded to spatial and spectral data of size m × n × 18 (where m×n is the image size in pixels) [42]. Furthermore, canonical discriminant analysis (CDA) was employed as a supervised transformation building method to divide the images into regions of interest (ROI) using the Videometer-Lab version 2.12.39 software (Videometer A/S, Herlev, Denmark). The final outcome of this segmentation process for each image was a dataset of spectral data including the average value and the standard deviation of the intensity of the pixels within the ROI at each wavelength.
FT-IR data were obtained using an FT-IR-6200 JASCO spectrometer (Jasco Corp., Tokyo, Japan) and a ZnSe 45 HATR (horizontal attenuated total reflectance) crystal (PIKE Technologies, Madison, WI, USA) with a refractive index of 2.4 and a depth of penetration of 2.0 µm at 1000 cm −1 . Spectra measurements were performed using Spectra Manager Code of Federal Regulations (CFR) software version 2 (Jasco Corp., Tokyo, Japan) in the wavenumber range of 4000-400 cm −1 , by accumulating 100 scans with a resolution of 4 cm −1 and a total integration time of 2 min.

Data Pre-Processing and Analysis
MSI spectral data were pre-processed by baseline offset treatment [43,44] for the development of PLS-R models in order to reduce random or systematic variations and simultaneously improve image resolution [45]. Likewise, for the development of the classification models, MSI data were subjected to standard normal variate (SNV) transformation prior to analysis [46]. Model training was undertaken with the dataset obtained from the storage experiments at isothermal conditions (n = 330), where 142 (43.1%) and 188 (56.9%) of the samples were defined as fresh (Class 1) and spoiled (Class 2), respectively. Model optimization was based on leave-one-out full-cross validation (LOOCV) process for PLS-R models and k-fold validation (k = 5) for the classification models. Moreover, the efficacy of the developed models to assess the quality of chicken samples was evaluated by external validation using independent datasets from the two dynamic temperature scenarios (n = 72; Class 1 = 36 samples, 50%; Class 2 = 36 samples, 50%).
FT-IR spectral data were modified by Savitzky-Golay first derivative (second polynomial order, 11-point window) for the development of PLS-R models, while for classification models' spectral data pre-treatment was based on the same model with a 9-point window in order to reduce baseline shift and noise [9]. Spectral data in the range of 1000 to 2000 cm −1 were included in the analysis, since these regions are documented as relevant to meat spoilage [37]. FT-IR models were also validated with data sets from dynamic temperature profiles (n = 63), including 30 (47.6%) fresh and 33 (52.4%) spoiled samples. The procedure of model training and validation is graphically presented in Figure 1.

Data Pre-Processing and Analysis
MSI spectral data were pre-processed by baseline offset treatment [43,44] for the development of PLS-R models in order to reduce random or systematic variations and simultaneously improve image resolution [45]. Likewise, for the development of the classification models, MSI data were subjected to standard normal variate (SNV) transformation prior to analysis [46]. Model training was undertaken with the dataset obtained from the storage experiments at isothermal conditions (n = 330), where 142 (43.1%) and 188 (56.9%) of the samples were defined as fresh (Class 1) and spoiled (Class 2), respectively. Model optimization was based on leave-one-out full-cross validation (LOOCV) process for PLS-R models and k-fold validation (k = 5) for the classification models. Moreover, the efficacy of the developed models to assess the quality of chicken samples was evaluated by external validation using independent datasets from the two dynamic temperature scenarios (n = 72; Class 1 = 36 samples, 50 %; Class 2 = 36 samples, 50%).
FT-IR spectral data were modified by Savitzky-Golay first derivative (second polynomial order, 11-point window) for the development of PLS-R models, while for classification models' spectral data pre-treatment was based on the same model with a 9-point window in order to reduce baseline shift and noise [9]. Spectral data in the range of 1000 to 2000 cm −1 were included in the analysis, since these regions are documented as relevant to meat spoilage [37]. FT-IR models were also validated with data sets from dynamic temperature profiles (n = 63), including 30 (47.6 %) fresh and 33 (52.4 %) spoiled samples. The procedure of model training and validation is graphically presented in Figure 1. PLS-R models for the estimation of TVCs and Pseudomonas spp. counts on chicken thighs surface were developed and validated by the software Unscrambler © ver. 9.7 (CAMO Software AS, Oslo, Norway). Moreover, linear discriminant analysis (LDA) [47], quadratic discriminant analysis (QDA) [31], support vector machines (SVMs), and quadratic support vector machines (QSVMs) [48] models were employed for the classification of samples according to their spoilage level using MATLAB 2012a software (The MathWorks, Inc., Natick, MA, USA). The performance of the developed models was PLS-R models for the estimation of TVCs and Pseudomonas spp. counts on chicken thighs surface were developed and validated by the software Unscrambler © ver. 9.7 (CAMO Software AS, Oslo, Norway). Moreover, linear discriminant analysis (LDA) [47], quadratic discriminant analysis (QDA) [31], support vector machines (SVMs), and quadratic support vector machines (QSVMs) [48] models were employed for the classification of samples according to their spoilage level using MATLAB 2012a software (The MathWorks, Inc., Natick, MA, USA). The performance of the developed models was evaluated via the following metrics and indexes: root mean squared error (RMSE), correlation coefficient (r), overall accuracy, sensitivity, and specificity [49,50].
The population dynamics of TVCs and Pseudomonas spp. on the surface of chicken thigh fillets stored at isothermal conditions (0, 5,10,15,20,25,30, and 35 °C) are presented in Figure 2. The initial population of TVC (Figure 2A,B) and Pseudomonas spp. (Figure 2C,D) was 4.02 (±0.38) and 3.75 (±0.11) log CFU/cm 2 , respectively, confirming previous literature findings [22,[51][52]. As expected, storage temperature significantly influenced microbial growth resulting in sample deterioration and spoilage. For poultry, TVCs values exceeding 7.0 log CFU/cm 2 have been reported by other researchers to signify the end of shelf-life due to spoilage [52][53][54]. More specifically, in this study TVCs reached values above 7.0 log CFU/cm 2 at 15 °C in 30 h (7.2 ± 0.15 log CFU/cm 2 ), at 10 °C in 72 h (7.24 ± 0.39 log CFU/cm 2 ), at 5 °C in 144 h (7.62 ± 0.63 log CFU/cm 2 ) and at 0 °C in 240 h (7.17 ± 0.42 log CFU/cm 2 ). Pseudomonas spp. counts were similar to TVCs population and spoilage was evident at 15 °C in 48 h (7.3 ± 0.33 log CFU/cm 2 ), at 10 °C in 72 h (7.06 ± 0.48 log CFU/cm 2 ), at 5 °C in 120 h (7.22 ± 0.18 log CFU/cm 2 ), and at 0 °C in 216 h (6.75 ± 0.23 log CFU/cm 2 ). Furthermore, samples appearance and odor rapidly deteriorated at high storage temperatures and TVCs reached 7.0 log CFU/cm 2 at 20 °C in 32 h (7.36 ± 0.39 log CFU/cm 2 ), at 25 °C in 24 h (7.78 ± 0.18 log CFU/cm 2 ), at 30 °C in 24 h (7.95 ± 0.40 log CFU/cm 2 ), and at 35 °C in 12 h (6.8 ± 0.46 log CFU/cm 2 ). Similarly, Pseudomonas spp. approached 7.0 log CFU/cm 2 at 20 °C in 32 h (6.97 ± 0.39 log CFU/cm 2 ), at 25 °C in 24 h (6.95 ± 0.36 log CFU/cm 2 ), at 30 °C in 24 h (6.89 ± 0.68 log CFU/cm 2 ), and at 35 °C in 24 h (6.69 ± 0.60 log CFU/cm 2 ). Moreover, the microbiological results from the two dynamic temperature profiles are shown in Figure 3. The initial TVCs and Pseudomonas spp. counts were 3.82 ± 0.21 log CFU/cm 2 and 2.51 ± 0.28 log CFU/cm 2 , respectively (first dynamic temperature profile, Figure 3A), and 4.13 ± 0.40 log CFU/cm 2 and 2.87 ± 0.65 log CFU/cm 2 , respectively (second Moreover, the microbiological results from the two dynamic temperature profiles are shown in Figure 3. The initial TVCs and Pseudomonas spp. counts were 3.82 ± 0.21 log CFU/cm 2 and 2.51 ± 0.28 log CFU/cm 2 , respectively (first dynamic temperature profile, Figure 3A), and 4.13 ± 0.40 log CFU/cm 2 and 2.87 ± 0.65 log CFU/cm 2 , respectively (second dynamic temperature profile, Figure 3B). Stored samples at these dynamic profiles were considered spoiled in 96 h (TVC = 6.96 ± 0.25 log CFU/cm 2 , Pseudomonas spp. = 6.19 ± 0.29 log CFU/cm 2 ) for the first dynamic profile and in 120 h (TVC = 7.08 ± 0.01 log CFU/cm 2 , Pseudomonas spp. = 7.05 ± 0.03 log CFU/cm 2 ) for the second dynamic profile. This one-day delay of spoilage could be attributed to the different metabolic footprint of chicken samples due to temperature alterations affecting thus microbial growth [55,56]. Statistical analysis for the microbiological results (one-way ANOVA via MATLAB 2012a software (The MathWorks, Inc., Natick, MA, USA)) is available in Tables S1 and S2. considered spoiled in 96 h (TVC = 6.96 ± 0.25 log CFU/cm 2 , Pseudomonas spp. = 6.19 ± 0.29 log CFU/cm 2 ) for the first dynamic profile and in 120 h (TVC = 7.08 ± 0.01 log CFU/cm 2 , Pseudomonas spp. = 7.05 ± 0.03 log CFU/cm 2 ) for the second dynamic profile. This one-day delay of spoilage could be attributed to the different metabolic footprint of chicken samples due to temperature alterations affecting thus microbial growth [55,56]. Statistical analysis for the microbiological results (one-way ANOVA via MATLAB 2012a software (The MathWorks, Inc., Natick, MA, USA)) is available in Table S1 and Table S2. More detailed information about chicken thigh fillets spoilage was derived by sensory evaluation, where 56.9% of the samples were scored above 2 and considered spoiled. Samples stored at 0 °C were considered acceptable until 240 h of storage, while samples stored at 30 and 35 °C were evaluated as spoiled after 6 and 12 h, respectively. In addition, deterioration of odor due to spoilage was evident in 96 h at 5 °C, 48 h at 10 °C, and 24 h at 15, 20, and 25 °C. The correlation of sensory scores to samples temperature and More detailed information about chicken thigh fillets spoilage was derived by sensory evaluation, where 56.9% of the samples were scored above 2 and considered spoiled. Samples stored at 0 • C were considered acceptable until 240 h of storage, while samples stored at 30 and 35 • C were evaluated as spoiled after 6 and 12 h, respectively. In addition, deterioration of odor due to spoilage was evident in 96 h at 5 • C, 48 h at 10 • C, and 24 h at 15, 20, and 25 • C. The correlation of sensory scores to samples temperature and TVCs populations is provided at Table 1. TVCs values above 6.99 log CFU/cm 2 corresponded to samples rated with an average score greater than 2, similarly to other studies where spoilage threshold was established at 7.0 log CFU/cm 2 for poultry [53]. Based on this criterion, samples were assigned in two quality classes, namely fresh (score < 2) or spoiled (score ≥ 2), and were further employed in the development of classification models.

Classification Models for the Assessment of Spoilage
The performance of the selected models to classify the samples in the respective quality classes (fresh or spoiled) through MSI spectral data is demonstrated by the confusion matrix (Table 4) for LDA, QDA, SVM, and QSVM. For the LDA model, 219 out of 330 samples and 49 out of 72 samples were correctly classified in both quality classes during model development (FCV) and prediction, respectively, providing overall accuracy of 66.4% and 68.1%. During FCV process, sensitivity and specificity were 59.4% and 73.3%, respectively, whereas for model prediction the calculated sensitivity and specificity were 76.0% and 63.8%, respectively. For QDA model, 214 out of 330 samples (overall accuracy 64.8%) and 50 out of 72 samples (overall accuracy 69.4%) were classified in the correct class for model FCV and prediction, respectively. Moreover, sensitivity and specificity were estimated at 57.6% and 72.8%, respectively, for model FCV and at 73.3% and 66.7%, respectively, for the model prediction. It is notable that improved results were obtained by the application of SVM model where 301 out of 330 samples (overall accuracy 91.2%) and 68 out of 72 samples (overall accuracy 94.4%) were correctly classified in the respective quality class during model development (FCV) and prediction, respectively. In addition, for SVM model sensitivity and specificity, percentages exhibited their highest values at 94.4% during external validation. Likewise, for QSVM implementation, 287 from 330 samples and 66 from 72 samples were efficiently identified during model development and prediction, with an overall accuracy of 87.0% and 91.7%, respectively. For this model, sensitivity and specificity percentages were calculated at 83.7% and 89.6% for model FCV, while for external validation the estimated values were 94.1% and 89.5%, respectively.

Classification Models for the Assessment of Spoilage
The performance of the selected models to classify the samples in the respective quality classes (fresh or spoiled) through MSI spectral data is demonstrated by the confusion matrix (Table 4) for LDA, QDA, SVM, and QSVM. For the LDA model, 219 out of 330 samples and 49 out of 72 samples were correctly classified in both quality classes during model development (FCV) and prediction, respectively, providing overall accuracy of 66.4% and 68.1%. During FCV process, sensitivity and specificity were 59.4% and 73.3%, respectively, whereas for model prediction the calculated sensitivity and specificity were 76.0% and 63.8%, respectively. For QDA model, 214 out of 330 samples (overall accuracy 64.8%) and 50 out of 72 samples (overall accuracy 69.4%) were classified in the correct class for model FCV and prediction, respectively. Moreover, sensitivity and specificity were estimated at 57.6% and 72.8%, respectively, for model FCV and at 73.3% and 66.7%, respectively, for the model prediction. It is notable that improved results were obtained by the application of SVM model where 301 out of 330 samples (overall accuracy 91.2%) and 68 out of 72 samples (overall accuracy 94.4%) were correctly classified in the respective quality class during model development (FCV) and prediction, respectively. In addition, for SVM model sensitivity and specificity, percentages exhibited their Regarding FT-IR classification models (  It needs to be noted that MSI-SVM and FT-IR-QDA combinations could not only efficiently classify samples in their correct quality class, with overall accuracy of 94.4% and 71.4%, respectively, but, simultaneously, the misclassified samples were equally distributed in the safe and in the dangerous side, with specificity reaching 94.4% and 72.7%, respectively. Another interesting finding from MSI-SVM model was the low difference in the overall accuracy percentages (91.2% vs. 94.4%) observed between model FCV and prediction, indicating robust model performance. Furthermore, the same trend was observed for sensitivity and specificity (94.0% in both cases). Previous researchers reported that SVMs could result in the development of robust regression and classification models for poultry products [31,58]. SVM and QSVM models were more suitable for MSI spectral data, with SVM linear classifiers presenting the best separation of data's hyperplane [31]. In contrast, probability parametric LDA and QDA models which assume that each class could be described as a multivariate normal distribution [29,31], exhibited better discrimination of classes for FT-IR data. This is in good agreement with other studies, where LDA was proposed as a supervised multivariate classification method in FT-IR spectroscopic analysis of meat samples [25]. Even though data matrices from MSI and especially FT-IR presented high dimensionality, there was no evident class imbalance according to the prediction performance of all developed models (Tables 4 and 5). Table 5. Confusion matrix and performance indexes of the developed classification models (LDA, QDA, SVM, QSVM) regarding sensory quality discrimination of chicken thigh samples based on FT-IR spectral data.

Conclusions
The findings of this research indicated that MSI spectral data combined with PLS-R could satisfactorily predict TVC and Pseudomonas spp. counts on the surface of chicken thigh fillets regardless of storage temperature and batch variation. Similarly, satisfactory performance was obtained by the implementation of the classification models, where MSI spectral data coupled to SVM model achieved the most accurate identification of quality classes (fresh or spoiled) among samples. Moreover, QDA was considered the most suitable classification model applied to FT-IR measurements. The generalization of the developed models was ascertained by the utilization of chicken samples stored at two dynamic temperature profiles simulating conditions of the distribution chain. Nevertheless, additional measurements and continuous information feedback (different seasons, packaging conditions, etc.) could improve the performance of the aforementioned models and thus result in more successful assessment of the quality of poultry products.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/10 .3390/foods10112723/s1, Table S1: One way ANOVA for the TVC of chicken thigh fillet samples at each isothermal storage temperature and dynamic temperature scenarios, Table S2: One way ANOVA for Pseudomonas spp. counts of chicken thigh fillet samples at each isothermal storage temperature and dynamic temperature scenarios.

Data Availability Statement:
The datasets generated for this study are available on request to the corresponding author.