Development of a predictive classification model and extraction of signature wavelengths for the identification of spoilage in chicken breast fillets during storage using Near Infrared Spectroscopy

doi:10.21203/rs.3.rs-4478852/v1

Download PDF

Research Article

Development of a predictive classification model and extraction of signature wavelengths for the identification of spoilage in chicken breast fillets during storage using Near Infrared Spectroscopy

https://doi.org/10.21203/rs.3.rs-4478852/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Technologies for rapid identification and prediction of food spoilage can be crucial in minimizing food waste and losses, although their efficiency requires further improvement. This study aimed to pinpoint specific near-infrared (NIR) wavelengths that could indicate spoilage in raw chicken breast fillets. In this study, commercial tray-packs of boneless, skinless chicken breast fillets stored in a walk-in cooler at 4˚C were periodically tested every other day until they reached the spoilage state (identified by > 7 log CFU/ml). A portable Hyper spectral spectroscopy device (Field Spec Hi-Res4), with a range of wavelengths of 350–2500 nm, was used to measure reflectance. In addition to hyper-spectral analysis, aerobic plate counts were conducted on the fillets. The data from these counts were then used to train a Back Propagation Neural Network (B.P.N.N.) with specific parameters (250,000 steps, a learning rate of 0.02, and 5 hidden layers) and Linear-Support Vector machines (SVM-Linear) with ten-fold cross-validation technique to categorize spoilage into three stages: baseline microbial count (up to 3 log CFU/ml) (Initiation), propagation (between 3 and 6.9 log CFU/ml), and spoiled (> 7 log CFU/ml). The feature extraction process successfully identified the most representative six signature wavelengths from the whole hyper-spectral profile, which facilitated the classification of different phases of spoilage. The BPNN model demonstrated a high classification accuracy, with 93.7% for baseline counts, 95.2% for the propagation phase, and 98% for the spoiled category. These signature hyperspectral wavelengths hold the potential for developing cost-effective and rapid food spoilage detection systems, particularly for perishable items.

Supervised Learning Algorithms

Andrew's curve analysis

Back Propagation Neural Network

Near-Infrared spectroscopy

The substantial expansion of consumer interest in chicken and poultry commodities over the past two decades and the increased worldwide import and export of these commodities have brought attention to reasonable indicators of the quality and integrity of food (King et al. 2017; Sharif and Zahid, 2018). Especially in growing and developing nations, where production opportunities have been considerably limited (Olsen & Hasan, 2012), poultry has emerged as the most demanded livestock commodity worldwide.

Leading global livestock producers such as Brazil, the United States, the European Union, and Thailand have boosted their domestic output in response to this growing demand (Narrod et al., 2011). Based on the survey data collected by Future Business insight (FBI) suggests that the meat and poultry industry in the United States is poised for substantial growth by 2028, particularly driven by the increasing popularity of nutrient-rich diets during post-COVID-19 pandemic (FBI, 2022). However, the burgeoning poultry industry is facing a significant impediment: food waste and losses (Kaur et al., 2023). With the United Nations expecting a significant expansion in the food market to support this development, driven by anticipated global population growth by 2050 (United Nations, 2021), addressing food waste becomes even more critical (Barrett, 2021). These losses occur across various stages of production, from post-harvest to consumer levels, in both developing and developed nations, underscoring the urgent need for effective management strategies (Gustavsson et al., 2011; Kitinoja et al., 2011).

Assessing the rate of microbial decomposition in raw poultry fillets during storage is a crucial step in addressing these challenges. Precise forecasting prediction models can help in improving preservation conditions, mitigate food waste, and ensure food safety (Tarlak, 2023). Strict control measures should be implemented throughout the entire food supply chain due to the presence of spoilage microorganisms that causes muscle degradation in fresh raw chicken products. The effectiveness of food safety management systems such as HACCP and ISO 22000 rely on accurate prediction and control of microbial growth (Biswas et al., 2019; Banerjee et al., 2019).

The breakdown of nutrient composition in poultry products due to spoilage emphasizes the need for optimal storage conditions to maintain quality and ensure food safety throughout the preservation process (Silva et al., 2018). Spoilage microbes can contaminate chicken products during harvesting, processing, handling, and packaging, highlighting the critical role of regulation in maintaining quality (Alexa et al., 2024). The United States faces challenges in preserving poultry products due to their perishable nature and the necessity to maintain quality and ensure food safety throughout extended marketing periods (Firouz et al., 2021). This requires specialized knowledge and safe handling at every step of the food security process to prevent waste and loss (Totobesola et al., 2022; Wang et al., 2021). Factors such as in-cell enzymes, nutrient availability, oxygen concentrations, and cross-contamination, along with temperature fluctuations, contribute to food spoilage (Adegoke, 2004). For fresh poultry commodities, modifications such as pH changes and fat oxidation can encourage microflora growth and metabolism, leading to the formation of toxic compounds and off-flavors (Sikorski et al., 2020; Huang et al., 2021; Herron et al., 2022).

Although fresh meat deterioration is a concern worldwide, more research is needed to understand the mechanisms and interactions causing spoilage (Herron et al., 2022). Several studies have investigated rapid detection techniques for spoilage in meat and related products. To address these challenges, various spectroscopy techniques have been explored for their potential in rapid spoilage detection (Soni et al. 2022). One of the frequently experimented techniques is the variations of spectroscopy that interact with sample material and changes in spectrum with different wavelengths. Methods such as reflectance spectroscopy, laser-induced breakdown spectroscopy, Raman spectroscopy, and fluorescence spectroscopy have shown promising results in assessing the quality of meat-based products and detecting microorganisms (Lin et al., 2004; Rehse, 2019; Kashif et al., 2021; Liu et al., 2020). However, the connection between microbial activity and metabolic spoilage, particularly concerning holding and shipping conditions, remains a subject requiring further exploration (in't Veld, 1996; Sikorski et al., 2020; Huang et al., 2021; Herron et al., 2022).

Laser-induced breakdown spectroscopy and Raman spectroscopy have been used for the rapid identification of different fish species, nuts adulteration, food authentications, and detection of adulterated powdered milk (Li et al., 2015; Ren et al., 2023; Sezer et al., 2019; Shin et al., 2023; Huang et al., 2022). Fluorescent spectroscopy is used for identifying different beverages, validating food authenticity, and evaluating quality validation in spices. Monitoring of shelf life and value-added food using mid-infrared spectroscopy is implemented in the food industry for classification and quality evaluation (Su & Sun, 2019). Over the past decade, researchers have explored the capabilities of instruments based on light emission interacting with surfaces based on their chemical and physical attributes, such as Hyper and Multispectral Imaging and vibrational spectroscopy like FT-IR (Bhargava, 2012; Candoğan et al., 2021). These techniques have shown promise in assessing the quality features of various foods and creating predictive models that gauge the quality and bacterial presence in numerous meat products (Shimoni & Labuza, 2000). Hyperspectral analysis has been used in the poultry sector to develop models that classify chicken breast fillets (Jiang et al., 2019). Using spectral data, some models can also determine bacterial levels in spoiling chicken meat (Ellis et al., 2002). FT-IR has shown efficiency in distinguishing fresh chicken breast muscles from spoiled ones, and its potential in detecting spoilage bacteria on chicken meat surfaces has been validated (Alexandrakis et al., 2012).

Given the increasing demands and challenges in the poultry sector regarding food quality and safety, the utilization of modern statistical and machine learning approaches has become crucial in overcoming these problems. Principal Component Analysis (PCA), a method with a rich history dating back to its introduction by Pearson in 1901 and further development by Hotelling in 1933, stands as a cornerstone in multivariate statistics (Sewell, 2008; Mishra et al., 2017). This historical context not only connects us to the pioneers of statistical analysis but also underscores the enduring relevance of PCA. It was designed to address the challenges posed by high-dimensional data; a concept first articulated by Bellman in 1961 (Bellman, 1961; Strange & Zwiggelaar, 2014). These challenges include overfitting, high computation costs, and the complexity of visualizing the data. PCA, a technique that transforms the original variables into uncorrelated variables known as principal components, effectively retains important changes in the data while reducing the number of dimensions. The first principal component captures the highest amount of variation, with each subsequent component collecting the remaining orthogonal variance (Jolliffe, 2002). PCA is a powerful technique that significantly reduces overfitting and processing demands, thereby enhancing model performance and data analysis. However, it's worth noting that PCA is sensitive to noise and outliers (Hubert et al., 2005).

Partial Least Squares (PLS) regression, a robust statistical technique, offers a unique approach to handling data with a high number of variables and situations involving collinearity (Abdi & Williams, 2013). This unique approach not only draws curiosity but also opens new possibilities in multispectral related data analysis. PLS, a statistical method that reduces dimensionality by combining principal component analysis and multiple regression, creates orthogonal latent variables that maximize covariance with the resultant variable (Abdi, 2010). This unique characteristic of PLS makes it particularly beneficial in disciplines such as chemometrics and bioinformatics (Jombart et al., 2010). While PLS is susceptible to overfitting and outliers (Wold et al., 2001), it provides valuable insights into complex connections and is essential for modeling sophisticated datasets.

Support Vector Regression (SVR), a Support Vector Machine (SVM) designed for regression problems, predicts continuous values by finding the optimal hyperplane in a high-dimensional space (Drucker et al., 1996). By using kernel functions, SVR effectively handles non-linear interactions, striking a balance between complexity and performance through careful hyperparameter tuning (Ben Ishak,2016; Smola & Schölkopf, 2004). The robustness of SVR in scenarios with different variables makes it well-suited for applications in finance, medicine, and energy. Despite its limitations due to computing limits and challenges in hyperparameter adaptation (Cao & Tay, 2003).

The Random Forest Regressor is an ensemble learning technique that uses many decision trees to provide precise and robust predictions (Bernard et al., 2009; Liaw & Wiener, 2015). This precision not only instills confidence in the model but also enhances the reliability of our predictions. Using randomness in data and feature selection alleviates overfitting, as Rodriguez-Galiano et al., 2015; Jun, 2021). This algorithm effectively manages extensive and intricate datasets, offering valuable insights into the important nature of features (Strobl et al., 2008). Nevertheless, the computational requirements and intricacies of hyperparameter tuning make it necessary for careful optimization (Liaw & Wiener, 2015).

The Gradient Boosting Regressor is a prediction model that achieves high accuracy by iteratively correcting errors using decision trees. This approach, which is efficient for various tasks, utilizes gradient descent to optimize the model by emphasizing residuals to enhance predictions (Nguyen et al., 2021; Awad & Khanna, 2015). Although it is prone to overfitting and requires significant computational resources, the model's effectiveness can be ensured through meticulous hyperparameter adjustment (Zhang & Haghani, 2015; Krauss et al., 2017).

The Multilayer Perceptron (MLP) is an artificial neural network that handles intricate classification and regression problems. The network consists of interconnected layers trained via backpropagation and gradient descent (Isabona et al., 2022; Badirli et al., 2020). Activation functions are used to introduce non-linearity. Although concerns are associated with overfitting, methods such as dropout and regularization can successfully alleviate these problems. The adaptability of MLP in machine learning is highlighted by its capacity to capture complex patterns, which necessitates meticulous optimization of network architecture and hyperparameters (Srivastava et al., 2014; Bergstra & Bengio, 2012).

This present study aims to identify specific hyperspectral reflectance wavelengths for developing rapid spoilage detection in fresh poultry during storage, along with feature extraction techniques and supervised machine learning algorithms. Additionally, this study provides fundamental information about the various analysis techniques that can be utilized for predictive modeling.

2.1. Data collection

Freshly deboned tray-packed breast filets from broilers (Ross 708) were transported into medium-sized isothermal coolers to the Auburn University Department of Poultry Science microbiological lab for the spoilage study. A total of 290 chicken breast fillets were used in the presented study. Breast fillets were randomly selected and stored at 4^oC until complete spoilage was reached. Standard aerobic plate count and hyperspectral spectra were collected at each storage period (0, 2, 4, 6 and 8 days). Stored breast fillets were subjected to hyperspectral spectroscopy for the collection of spectral data using a handheld Field Spec Hi-Rres 4 device (Pan analytical, MI, United States), equipped with spectral data processing software ViewSpecPro (developed by RJL Systems, Detroit, MI, United States). A handheld probe was placed around 1 cm above the ventral surface of the breast fillet. Using ViewSpecPro software, six measurements were taken on the breast fillets and averaged for the mean value for reflectance at different wavelengths (Fig. 2). Microbiological analysis of the stored chicken breast fillets were performed using the protocol mentioned in Herron et al. (2022). Different supervised ML regression models were also developed and compared for the prediction of logCFU/ml value.

2.2. Data processing and screening

2.2.1. Predictor Screening and Cluster Variable Analysis

The collected data was analyzed using J.M.P. Pro 17 software for predictor screening and cluster variable analysis. Predictor screening was performed based on the bootstrap forest method with 100 randomly generated trees with a set seed value of 2000. Andrew curve analysis was performed on JASP. Software for preprocessed or uncleaned data and post-processed or tidy data. Data that tends to form clusters shows a fine curve when transformed using the Fourier transformation equation. Specific hyperspectral wavelengths were selected based on the ranking decided by the software itself using contribution and distributed ranks. Selected top 200 hyperspectral wavelengths from the predictor screening process were subjected to cluster variable analysis for the secondary screening using the labeled column (Initiation, Propagation, and spoiled). Cluster variable analysis is less popular in statistics and refers to the method where variables are clustered instead of individual observations based on their pattern or relationship (Anderberg, 2014). The selected parameters then would be used as inputs in various predictive analysis and classification approaches, such as support vector machines and neural networks for developing models. It is rare to find a detailed explanation of the variable clustering algorithm method in most textbooks that cover multivariate approaches. However, the variable cluster tracks progress primarily as an application in statistical analysis software. Most representative hyperspectral wavelengths observed in variable clustering analysis were selected to develop classification and predictive models using a Back Propagation Neural Network and Support Vector Machine.

2.2.2. Andrew Curve Analysis

The Andrew curve or plot visualizes multivariate data by mapping each observation into a function (t). David F Andrew developed the method to visualize high-dimensional data into individual observation curves. As shown below, if given q is the given variable for each observation, these variables were considered as co-efficient in a mathematical equation for Fourier series as:

f(t) = x0+x1sin(t) + x2cos(t) + x3sin(2t) + x4cos(2t) +…………………… (1)

In the above equation, t is the expansion variable, typically ranging from -π to π, and xi denotes the variable of the given observation from the dataset. Andrew curve analysis provides several benefits, such as a reduction in high dimensionality, which is unnecessary as it can deal with any given dimensions and helps in pattern recognition in the datasets. After the collected and preprocessed data was downloaded for the development of the predictive model, data was subjected to Andrew Curve analysis to visualize the structure of collected preprocessed data and post-processed data, such as overlapping of data, and to make the estimation of further preprocessing of data to avoid underfitting and overfitting during model development. Estimation was made before and after data processing based on predictor screening results (Fig. 3A & B).

The PCA analysis was performed on chicken breast fillets maintained at a temperature of 4°C until spoiled, using Near-Infrared (NIR) wavelengths ranging from 350 nm to 2500 nm. This analysis revealed significant patterns and important insights that are crucial for comprehending the spectrum variations and their associated implications. The first principal component (PC1) exhibited a significant influence, capturing 96.4% of the total variation. The observed strong dominance implies the presence of a substantial underlying component or a combination of factors that primarily influence the hyperspectral characteristics of the chicken breast samples in the analyzed conditions. The extent of variation represented by PC1 may suggest changes in moisture content, protein configurations, or adjustments in fat composition, which are well-known factors that significantly affect the near-infrared spectra in fresh poultry products. One notable finding was the dense grouping of most of these wavelengths, indicating a significant level of uniformity within the given storage conditions. These observed patterns may suggest that the storage conditions provide a stable and uniform environment, resulting in limited fluctuations in the hyperspectral signature of chicken breast fillets. Moreover, the identification of these wavelengths may exhibit a direct correlation with distinct chemical or physical features that are intrinsic to chicken breast fillets. This correlation might potentially serve as markers or indicators for specific attributes or quality factors. Several studies have found that the hyperspectral data variables were explained by first principal components with a varying range of 36–88% of variance (Sahar et al. 2018; Alexandrakis et al. 2012).

Results obtained from development of regression models shown in Table 1, The Random Forest Regression model demonstrates superior performance in terms of Mean Squared Error (MSE), exhibiting an average MSE of 0.17. This outcome suggests a high level of precision and accuracy in the model's predictive capabilities. The model's consistency is emphasized by a low standard deviation of 0.15 for mean squared error (MSE), indicating that it consistently performs adequately across dataset and towards different folds (In this study 10-fold) of cross-validation. In contrast, Gradient Boosting exhibits robust accuracy, as evidenced by its mean squared error (MSE) of 0.19 and standard deviation of 0.16. Additionally, it demonstrates a marginally higher level of variability compared to Random Forest.

Table 1

Summary table for the comparison of different machine learning based regression models.
Model Name	Mean Squared Error Mean	Mean Squared Error std	R-Squared Mean	R-squared std	Mean Absolute Error Mean	Mean Absolute Error Mean std
SVM Regression	0.25	0.32	0.930	0.08	0.23	0.11
Random Forest Regression	0.17	0.15	0.936	0.04	0.17	0.04
Gradient Boosting	0.19	0.16	0.954	0.04	0.18	0.05
Neural Network Regression	0.26	0.13	0.936	0.03	0.36	0.06
Partial Least Square Regression	0.47	0.15	0.888	0.03	0.53	0.08

When evaluating the model's capacity to account for the variability in the response data, Gradient Boosting demonstrates a significant value of 0.954 for R². The previously stated values clearly indicates that Gradient Boosting has the capability to explain about 95.4% of the variability observed in the dependent variable. Both the Random Forest and Neural Network Regression models exhibit a mean R² value of 0.936, indicating a robust ability to forecast variance. However, it is worth noting that their predictive power is slightly inferior to that of the Gradient Boosting model.

The Random Forest Regression model accuracy results are represented in the terms of Mean Absolute Error (MAE), as indicated by its lowest mean MAE value of 0.17. This finding further supports the model's capability to generate predictions that closely align with the actual values. Gradient Boosting demonstrates an average Mean Absolute Error (MAE) of 0.18, which exhibits a level of efficiency equivalent to that of Random Forest model. In contrast, it is observed that Neural Network Regression has the highest mean MAE value of 0.36, indicating that its average predictions have comparatively lower accuracy when compared to the other models. However, the Partial Least Square Regression has shown lower performance across all metrics, with a mean MSE of 0.47 being the greatest, a mean R² of 0.888 being the lowest, and a mean MAE of 0.53 being the highest. Based on the presented data, it can be inferred that this particular model exhibits a somewhat lower level of accuracy and effectiveness in elucidating the variability in the dependent variable, as compared to the other models under consideration.

Andrew curve analysis for the unprocessed data shows reflectance values were overlapped on each wavelength from 350–2500 nm (Fig. 3A). The post-processing of collected data using the bootstrapping method's predictor analysis revealed clear, less overlapped reflectance curves (Fig. 3B). 200 wavelengths were selected based on the contribution and ranking from 1-200. Different supervised machine learning models, such as SVM and BPNN, were developed and compared for testing accuracy. To avoid underfitting and overfitting, the model was validated using a ten times k-fold cross-validation technique in jmp17 pro software and R software (Version 4.0.3; Bunny-Wunnies Freak Out). Feature extraction of the collected data using predictor screening and variable cluster analysis resulted in different wavelengths for each spoilage phase. Wavelengths of 385 nm and 2292 nm were clustered for the initiation phase, Wavelengths of 432 nm and 2311 nm were clustered for the propagation phase, and wavelengths of 1141 nm and 2305 nm were clustered for the spoilage phase of the analysis. Wavelengths clustered for initiation, propagation, and spoiled phase represented 95%, 91.1%, and 93.3% of the variation proportion, indicating that grouping of data using clustering captures its inherent variability, respectively (Table 2).

Table 2

Summary of Cluster Variable analysis for feature extraction process in collected NIR dataset.
Spoilage Phase	Wavelengths	Cluster proportion of Variation explained.	Average Clustering proportion of Variation	1-R² Ratio
Initiation	385 nm	96.2%	94.9%	0.07
	2292 nm	93.7%	94.9%	0.02
Propagation	432 nm	90.3%	91.1%	0.06
	2305 nm	92.1%	91.1%	0.02
Spoiled	1141 nm	95.0%	93.3%	0.01
	2292 nm	91.3%	93.3%	0.02

A study conducted by Lin et al. 2004 evaluated the use of Visible and short-wave NIR (SW-NIR) to detect and quantify the microbial loads on chicken breast muscles. Authors have used the full spectra wavelengths from 600–1100 nm for the study and implemented PCA. along with partial least square (PLS) based predictions and found that PCA analysis was able to separate the samples for eight hours as compared to 0-hrs with eight latent variables with an R-value of 0.91. Another study conducted by Alexandrakis (2012) on non-destructive rapid chicken breast spoilage identification using NIR spectra and analysis through PCA showed that this analysis of the MSC-modified spectral data between wavelengths 1,550–1,996 nm could not entirely differentiate spoilage conditions at day 0 from day four samples of chicken muscles. Score plots for days 0 and 8 showed little overlap, and those for days 0 and 14 were distinct in the PC1-PC2 space. Also, as the storage duration increases, spoilage begins, leading to muscle protein breakdown, causing more noticeable alterations in the muscle's spectral response, which indicates clear separation observed in days 0, 8, and 14 samples. By the 14th day, the changes are more intricate, hinting that the NIR spectra capture multiple phenomena. However, their findings suggest that NIR spectroscopy could identify chicken breast muscle spoilage when the microbial count on chicken fillets exceeds 6.75 log10 cfu g − 1, the average count for day four samples not at different spoilage phases.

Jiang et al. 2021, assessed high-resolution spectral data between the wavelength of 900–1700 nm. They have used techniques like Savitzky-Golay convolution smoothing (S.G.C.S.), standard normal variate (S.N.V.), and multiplicative scatter correction (M.S.C.) for data post-processing steps. Subsequently, they employed the partial least squares (P.L.S.) algorithm to establish a relationship with the total quantities of Pseudomonas spp. and Enterobacteriaceae (PEC) in fresh chicken breasts, aiming for a swift PEC prediction. Findings from their study revealed that the MSC-PLS model, using the entire 900–1700 nm range and MSC spectra, outperformed other P.L.S. spectral models, achieving a correlation coefficient of 0.954 and a prediction error of 0.396 log10 CFU/g. When focusing on 12 prime wavelengths (902.2 nm, 905.5 nm, 923.6 nm, 938.4 nm, 946.7 nm, 1025.7 nm, 1124.4 nm, 1211.6 nm, 1269.2 nm, 1653.7 nm, 1691.8 nm and 1693 nm) identified from M.S.C. spectra using the successive projections algorithm (SPA), the SPA-MSC-PLS model yielded results comparable to the MSC-PLS model.

The performance of the Back Propagation Neural Network (BPNN) model (a supervised machine learning model) in predicting the spoilage condition (Initiation, Propagation, and Spoiled) of raw poultry meat during the experimental time was evaluated. The model was developed using a 10-fold cross-validation technique with the nested 5-fold cross-validation method to deal with an unbalanced post-processed dataset with a set seed value of 2000. The model's performance is assessed using various statistical measures, such as Generalized R2, Entropy R2, RASE (Relative Approximation Squared Error), misclassification rate, and Receiver Operating Characteristics (R.O.C.). The B.P.N.N. model demonstrated a strong performance, explaining 96.1% of the variance in the data with a Generalized R² value of 0.961 in the testing data set with 10-fold cross-validation combined with a 5-fold nested cross-validation procedure. Entropy R² of 0.889 suggests a better ability to predict the spoilage phase during storage. A RASE value of 0.18 denotes a relatively low error rate, and a misclassification Rate of 0.025 implies that 2.5% of the model's predictions are incorrect. The testing accuracy for the different spoilage phase conditions ranged from 95.2–100%. Classification accuracy for testing data results showed 93.7% accuracy for the initiation phase, 95.2% accuracy for propagation, and 98% accuracy for spoiled raw poultry fillets (Fig. 4).

The results obtained from the SVM classification algorithm showed that the percent accuracy for the classification of the initiation phase in the test data set was 87.07%; for the propagation phase of the spoiled condition, the testing set showed an accuracy of 84.75%; and the classification accuracy for the spoiled phase of raw chicken fillets for the test data set was 64.58% (Fig. 4). Due to the uneven random distribution of data points in the training and testing split method, we also performed the 10-fold cross-validation technique to determine the overall accuracy of the developed SVM model for collected post-processed NIR wavelength spoilage datasets. Our results indicated that the cost function of C = 1.5, with repeated cross-validation accuracy, for the developed model was 75.9% (Fig. 5).

Various statistical metrics were considered in evaluating the classification model's performance. The developed SVM model demonstrated an overall accuracy of 75.9%, indicating a relatively high level of correct predictions across all categories. The 95% confidence interval for the accuracy ranged from 71.62–85.67%, assuring that the true accuracy of the model is likely within this range. Compared to the No Information Rate (NIR) of 52.86%, the model's accuracy is significantly higher, as supported by a p-value less than 0.05. This suggests that the model's predictions are informative and do not occur by random chance. The study conducted by Spyrelli et al. (2021) used nine different machine learning algorithms using SorfML software, which includes SVM Linear, SVM radial basis function, random forest, k-Nearest Neighbor, Neural Network (nnet), ridge, lars, pcr, and pls for the data collected on the spoilage of chicken breast fillets using Fourier's transformed infrared spectroscopy (FT-IR), and multispectral image (MSI) analysis with the range of wavelength from 405–970 nm. The authors have found that the highest performance was given by the neural network model with an average accuracy of 81.6%, followed by the SVM linear model with an average accuracy of 75.8%, and radial basis function SVM algorithms with an accuracy of 56.21%.

Overall, this study has shown that UV-Vis Hyperspectral spectroscopy, when combined with feature extraction techniques and supervised machine learning techniques, has great potential for quickly detecting and identifying spoilage in raw poultry at various stages of storage and spoilage. The findings of feature extraction, obtained from predictor analysis and variable clustering, have accurately identified the six most representative wavelengths from the whole set of hyperspectral bands. Creating hyperspectral imaging systems that are calibrated to these wavelengths can result in enhanced accuracy when identifying spoilage. These customized instruments can analyze the unique patterns that are associated with decay, thereby enabling continuous monitoring and early identification of quality reduction.

The classification results provided by the BPNN and SVM models provide opportunities for using supervised machine learning techniques in extensive studies. The backpropagation neural network model demonstrated improved ability in differentiating between different phases of spoiling compared to the SVM-linear model. The implementation of near-infrared spectroscopy (NIR), along with these machine learning (ML) approaches, suggests a future where the monitoring of bacterial development in poultry products becomes a standard practice, establishing itself as a reliable indicator for ensuring food quality.

Furthermore, there is a potential for the creation of sophisticated hyperspectral sensors that can specifically target these wavelengths to capture precise and particular information related to spoilage. This technological advancement would enable us to not only identify spoilage, but also analyze the chemical compositions and detect the presence of microorganisms that indicate spoilage. These improvements would help us improve our understanding of how food spoils and the resulting changes in taste, smell, and texture. This would allow us to take a comprehensive approach to managing food safety. Hence, additional investigation is crucial not only to determine the complex connection between NIR wavelengths and the chemical indicators of spoilage but also to comprehend the specific patterns of deterioration. This knowledge will aid in developing comprehensive strategies for managing spoilage.

Authors Contribution

Conceptualization, Amit Morey; Data curation, Aftab Siddique, Charles B Herron, Bet Wu, Katherine S.S. Melendrez, and Luis J. G. Sabillon; Formal analysis, Aftab Siddique; Funding acquisition, Amit Morey; Investigation, Aftab Siddique and Amit Morey; Methodology, Aftab Siddique, and Amit Morey; Writing—original draft, Aftab Siddique; Writing—review & editing, Laura J. Garner, Mary Durstock, and Alvaro Sanz-Saez; Equipment and resources, , Mary Durstock, and Alvaro Sanz-Saez. All authors have read and agreed to the published version of the manuscript.

Funding

The Present work was supported by FFAR new innovator food and agriculture research.

Conflict of Interest

The authors declare no conflicts of interest.

Acknowledgments

The authors would like to acknowledge the Foundation for Food and Agriculture Research for providing funds to conduct the research, the Department of Poultry Sciences for providing support, the Department of Crop Soils and Environmental Sciences for providing a UV-Vis Hyperspectral spectrometer for this study, and our research collaborators for providing chicken breast fillets for the study.

Abdi, H. (2010). Partial least squares regression and projection on latent structure regression (PLS Regression). Wiley interdisciplinary reviews: computational statistics, 2(1), 97-106.
Abdi, H., & Williams, L. J. (2013). Partial least squares methods: partial least squares correlation and partial least square regression. Computational Toxicology: Volume II, 549-579.
Adegoke, G.O., 2004. Understanding food microbiology.
Alexa, E. A., Papadochristopoulos, A., O’Brien, T., & Burgess, C. M. (2024). Microbial contamination of food. In Food Packaging and Preservation (pp. 3-19). Academic Press.
Alexandrakis, D., Downey, G. and Scannell, A.G.M., 2012. Rapid non-destructive detection of spoilage of intact chicken breast muscle using near-infrared and Fourier transform mid infrared spectroscopy and multivariate statistics. Food and Bioprocess Technology, 5(1), pp.338-347.
Anderberg, M.R., 2014. Cluster analysis for applications: probability and mathematical statistics: a series of monographs and textbooks Vol. 19. Academic Press.
Awad, M. and Khanna, R., 2015. Efficient learning machines: theories, concepts, and applications for engineers and system designers. Springer Nature.
Badirli, S., Liu, X., Xing, Z., Bhowmik, A., Doan, K. and Keerthi, S.S., 2020. Gradient boosting neural networks: Grownet. arXiv preprint arXiv:2002.07971.
Banerjee, D. K., Das, A. K., Thakur, N., Talukder, S., Das, A., Sonowal, J., ... & Sakunde, D. T. (2019). Factors affecting microbial growth in livestock products: A review. Ijcs, 7(3), 4017-4022.
Barrett, C. B. (2021). Overcoming global food security challenges through science and solidarity. American Journal of Agricultural Economics, 103(2), 422-447.
Bellman, R., 1961. Adaptive Control Processes: A Guided Tour. Princeton University Press.
Ben Ishak, A., 2016. Variable selection using support vector regression and random forests: A comparative study. Intelligent Data Analysis, 20(1), pp.83-104.
Bergstra, J. and Bengio, Y., 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(Feb), pp.281-305.
Bernard, S., Heutte, L. and Adam, S., 2009. On the selection of decision trees in random forests. In 2009 International Joint Conference on Neural Networks (pp. 302-307). IEEE.
Bhargava, R. (2012). Infrared spectroscopic imaging: the next generation. Applied spectroscopy, 66(10), 1091-1120.
Biswas, C., Leboveic, A., Burke, K. and Biswas, D., 2019. Postharvest approaches to improve poultry meat safety. In Food Safety in Poultry Meat Production (pp. 123-138). Springer, Cham.
Candoğan, K., Altuntas, E. G., & İğci, N. (2021). Authentication and quality assessment of meat products by fourier-transform infrared (FTIR) spectroscopy. Food Engineering Reviews, 13, 66-91.
Cao, L.J. and Tay, F.E.H., 2003. Support vector machine with adaptive parameters in financial time series forecasting. IEEE Transactions on Neural Networks, 14(6), pp.1506-1518.
Das, A.K., Nanda, P.K., Das, A. and Biswas, S., 2019. Hazards and safety issues of meat and meat products. In Food Safety and Human Health (pp. 145-168). Academic Press.
Drucker, H., Burges, C.J., Kaufman, L., Smola, A. and Vapnik, V., 1996. Support vector regression machines. In: Advances in neural information processing systems, 9.
Ellis, D. I., Broadhurst, D., Kell, D. B., Rowland, J. J., & Goodacre, R. (2002). Rapid and quantitative detection of the microbial spoilage of meat by Fourier transform infrared spectroscopy and machine learning. Applied and environmental microbiology, 68(6), 2822-2828.
Feng, Y.Z. and Sun, D.W., 2013a. Determination of total viable count (TVC) in chicken breast fillets by near-infrared hyperspectral imaging and spectroscopic transforms. Talanta, 105, pp.244-249.
Feng, Y.Z. and Sun, D.W., 2013b. Near-infrared hyperspectral imaging in tandem with partial least squares regression and genetic algorithm for non-destructive determination and visualization of Pseudomonas loads in chicken fillets. Talanta, 109, pp.74-83.
Feng, Y.Z., ElMasry, G., Sun, D.W., Scannell, A.G., Walsh, D. and Morcy, N., 2013. Near-infrared hyperspectral imaging and partial least squares regression for rapid and reagentless determination of Enterobacteriaceae on chicken fillets. Food Chemistry, 138(2-3), pp.1829-1836.
Firouz, M.S., Mohi-Alden, K. and Omid, M., 2021. A critical review on intelligent and active packaging in the food industry: Research and development. Food Research International, 141, 110113.
Fortune Business Insight, 2022. U.S. meat market. Available at: https://www.fortunebusinessinsights.com/u-s-meat-market-105342 [Accessed 18 June 2022].
Gustavsson, J., Cederberg, C., Sonesson, U., van Otterdijk, R. & Meybeck, A. 2011, Global food losses and food waste: extent, causes, and prevention, FAO, Rome.
Herron, C.B., Garner, L.J., Siddique, A., Huang, T.S., Campbell, J.C., Rao, S. & Morey, A. 2022, 'Building “First Expire, First Out” Models to Predict Food Losses at Retail due to Cold Chain Disruption in the Last Mile', Frontiers in Sustainable Food Systems, p. 471.
Hotelling, H. 1933, 'Analysis of a complex of statistical variables into principal components', Journal of Educational Psychology, vol. 24, no. 6, pp. 417-441.
Huang, J., Zhou, Y., Chen, M., Huang, J., Li, Y. & Hu, Y. 2021, 'Evaluation of negative behaviors for single specific spoilage microorganism on little yellow croaker under modified atmosphere packaging: Biochemical properties characterization and spoilage related volatiles identification', LWT, p. 140, 110741.
Huang, W., Guo, L., Kou, W., Zhang, D., Hu, Z., Chen, F., ... & Cheng, W. 2022, 'Identification of adulterated milk powder based on convolutional neural network and laser-induced breakdown spectroscopy', Microchemical Journal, vol. 176, 107190.
Hubert, M., Rousseeuw, P. J., & Vanden Branden, K. (2005). ROBPCA: a new approach to robust principal component analysis. Technometrics, 47(1), 64-79.
Illikoud, N., Rossero, A., Chauvet, R., Courcoux, P., Pilet, M.F., Charrier, T., ... & Zagorec, M. 2019, 'Genotypic and phenotypic characterization of the food spoilage bacterium Brochothrix thermosphacta', Food Microbiology, vol. 81, pp. 22-31.
in't Veld, J.H.H. 1996, 'Microbial and biochemical spoilage of foods: an overview', International Journal of Food Microbiology, vol. 33, no. 1, pp. 1-18.
Isabona, J., Imoize, A.L., Ojo, S., Karunwi, O., Kim, Y., Lee, C.C. & Li, C.T. 2022, 'Development of a multilayer perceptron neural network for optimal predictive modeling in urban microcellular radio environments', Applied Sciences, vol. 12, no. 11, 5713.
Jiang, H., Yoon, S. C., Zhuang, H., Wang, W., Li, Y., & Yang, Y. (2019). Integration of spectral and textural features of visible and near-infrared hyperspectral imaging for differentiating between normal and white striping broiler breast meat. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 213, 118-126.
Jolliffe, I.T. 2002, Principal component analysis for special types of data, Springer, New York.
Jombart, T., Devillard, S., & Balloux, F. (2010). Discriminant analysis of principal components: a new method for the analysis of genetically structured populations. BMC genetics, 11, 1-15.
Jun, M.J. 2021, 'A comparison of a gradient boosting decision tree, random forests, and artificial neural networks to model urban land use changes: The case of the Seoul metropolitan area', International Journal of Geographical Information Science, vol. 35, no. 11, pp. 2149-2167.
Kashif, M., Majeed, M. I., Nawaz, H., Rashid, N., Abubakar, M., Ahmad, S., ... & Anwar, M. A. (2021). Surface-enhanced Raman spectroscopy for identification of food processing bacteria. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 261, 119989.
Kaur, M., Singh, A. K., & Singh, A. (2023). Bioconversion of food industry waste to value added products: Current technological trends and prospects. Food Bioscience, 102935.
Kharola, S., Ram, M., Mangla, S.K., Goyal, N., Nautiyal, O.P., Pant, D. & Kazancoglu, Y. 2022, 'Exploring the green waste management problem in food supply chains: A circular economy context', Journal of Cleaner Production, vol. 351, 131355.
King, T., Cole, M., Farber, J. M., Eisenbrand, G., Zabaras, D., Fox, E. M., & Hill, J. P. (2017). Food safety for food security: Relationship between global megatrends and developments in food safety. Trends in Food Science & Technology, 68, 160-175.
Kitinoja, L., Saran, S., Roy, S. K., & Kader, A. A. (2011). Postharvest technology for developing countries: challenges and opportunities in research, outreach and advocacy. Journal of the Science of Food and Agriculture, 91(4), 597-603.
Krauss, C., Do, X.A. and Huck, N., 2017. Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500. European Journal of Operational Research, 259(2), pp.689-702.
Li, S., Li, L., Zeng, Q., Zhang, Y., Guo, Z., Liu, Z., ... & Liu, S. (2015). Characterization and noninvasive diagnosis of bladder cancer with serum surface enhanced Raman spectroscopy and genetic algorithms. Scientific reports, 5(1), 9582.
Liaw, A. and Wiener, M., 2015. randomForest: Breiman and Cutler’s random forests for classification and regression. R package version, 4, p.14.
Lin, M., Al-Holy, M., Mousavi-Hesary, M., Al-Qadiri, H., Cavinato, A.G. and Rasco, B.A., 2004. Rapid and quantitative detection of microbial spoilage in chicken meat by diffuse reflectance spectroscopy (600–1100 nm). Letters in Applied Microbiology, 39(2), pp.148-155.
Lins, M., Puppin Zandonadi, R., Raposo, A. and Ginani, V.C., 2021. Food waste in food service: an overview through the perspective of sustainable dimensions. Foods, 10(6), 1175.
Liu, H., Ji, Z., Liu, X., Shi, C., & Yang, X. (2020). Non-destructive determination of chemical and microbial spoilage indicators of beef for freshness evaluation using front-face synchronous fluorescence spectroscopy. Food chemistry, 321, 126628.
Liu, X. R., Zhang, M. M., & Gross, M. L. (2020). Mass spectrometry-based protein footprinting for higher-order structure analysis: fundamentals and applications. Chemical reviews, 120(10), 4355-4454.
Liu, Y., Sun, D.W., Cheng, J.H. and Han, Z., 2018. Hyperspectral imaging sensing of changes in moisture content and color of beef during microwave heating process. Food Analytical Methods, 11, pp.2472-2484.
Mishra, S. P., Sarkar, U., Taraphder, S., Datta, S., Swain, D., Saikhom, R., ... & Laishram, M. (2017). Multivariate statistical data analysis-principal component analysis (PCA). International Journal of Livestock Research, 7(5), 60-78.
Narrod, C., Tiongco, M., & Scott, R. (2011). Current and predicted trends in the production, consumption and trade of live animals and their products. Rev. sci. tech. Off. int. Epiz, 30(1), 31-49.
Natekin, A. and Knoll, A., 2013. Gradient boosting machines, a tutorial. Frontiers in Neurorobotics, 7, p.21.
National Chicken Council, 2022. Per Capita Consumption of Poultry and Livestock, 1965 to Forecast 2022, in Pounds. [online] Available at: https://www.nationalchickencouncil.org/about-the-industry/statistics/per-capita-consumption-of-poultry-and-livestock-1965-to-estimated-2012-in-pounds/ [Accessed 18 June 2022].
Nguyen, D. H., Le, X. H., Heo, J. Y., & Bae, D. H. (2021). Development of an extreme gradient boosting model integrated with evolutionary algorithms for hourly water level prediction. IEEE Access, 9, 125853-125867.
Olsen, R. L., & Hasan, M. R. (2012). A limited supply of fishmeal: Impact on future increases in global aquaculture production. Trends in Food Science & Technology, 27(2), 120-128.
Olsen, R.L. and Hasan, M.R., 2012. A limited supply of fishmeal: Impact on future increases in global aquaculture production. Trends in Food Science & Technology, 27(2), pp.120-128.
Pearson, K., 1901. LIII. On lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11), pp.559-572.
Rehse, S. J. (2019). A review of the use of laser-induced breakdown spectroscopy for bacterial classification, quantification, and identification. Spectrochimica Acta Part B: Atomic Spectroscopy, 154, 50-69.
Ren, L., Tian, Y., Yang, X., Wang, Q., Wang, L., Geng, X. and Lin, H., 2023. Rapid identification of fish species by laser-induced breakdown spectroscopy and Raman spectroscopy coupled with machine learning methods. Food Chemistry, 400, p.134043.
Ren, X., Zhao, J., Li, X., Shao, J., Pan, B., Salamé, A., ... & Liu, B. (2023). In-situ spectroscopic probe of the intrinsic structure feature of single-atom center in electrochemical CO/CO2 reduction to methanol. Nature Communications, 14(1), 3401.
Sahar, A., Pasha, I. and Ishaq, A., 2018. Assessing the capability of Fourier transform infrared spectroscopy in tandem with chemometric analysis for predicting poultry meat spoilage. PeerJ.
Schölkopf, B., Smola, A. and Müller, K.R., 1998. Nonlinear component analysis as a kernel eigenvalue problem. Neural Computation, 10(5), pp.1299-1319.
Sewell, M. (2008). Principal component analysis. University College London: London, UK.
Sezer, B., Apaydin, H., Bilge, G. and Boyaci, I.H., 2019. Detection of Pistacia vera adulteration by using laser induced breakdown spectroscopy. Journal of the Science of Food and Agriculture, 99(5), pp.2236-2242.
Sharif, M. K., & Zahid, A. (2018). Role of food product development in increased food consumption and value addition. In Food processing for increased quality and consumption (pp. 455-479). Academic Press.
Sharif, M. K., & Zahid, A. (2018). Role of food product development in increased food consumption and value addition. In Food processing for increased quality and consumption (pp. 455-479). Academic Press.
Shimoni, E. and Labuza, T.P., 2000. Modeling pathogen growth in meat products: future challenges. Trends in Food Science & Technology, 11(11), pp.394-402.
Shin, H., Kang, Y., Choi, K. W., Kim, S., Ham, B. J., & Choi, Y. (2023). Artificial intelligence-based major depressive disorder (MDD) diagnosis using raman spectroscopic features of plasma exosomes. Analytical chemistry, 95(15), 6410-6416.
Shin, S., Doh, I.J., Okeyo, K., Bae, E., Robinson, J.P. and Rajwa, B., 2023. Hybrid Raman and Laser-Induced Breakdown Spectroscopy for Food Authentication Applications. Molecules, 28(16), p.6087.
Sikorski, Z.E., Kołakowska, A. and Burt, J.R., 2020. Postharvest biochemical and microbial changes. In: Seafood: Resources, nutritional composition, and preservation, 55-75. Boca Raton, FL: CRC Press.
Silva, F., Domingues, F. C., & Nerín, C. (2018). Trends in microbial control techniques for poultry products. Critical reviews in food science and nutrition, 58(4), 591-609.
Smola, A.J. and Schölkopf, B., 2004. A tutorial on support vector regression. Statistics and Computing, 14(3), pp.199-222.
Soni, A., Dixit, Y., Reis, M. M., & Brightwell, G. (2022). Hyperspectral imaging and machine learning in food microbiology: Developments and challenges in detection of bacterial, fungal, and viral contaminants. Comprehensive Reviews in Food Science and Food Safety, 21(4), 3717-3745.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. and Salakhutdinov, R., 2014. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1), pp.1929-1958.
Strange, H., & Zwiggelaar, R. (2014). Open problems in spectral dimensionality reduction.
Strobl, C., Boulesteix, A.L., Kneib, T., Augustin, T. and Zeileis, A., 2008. Conditional variable importance for random forests. BMC Bioinformatics, 9, pp.1-11.
Su, W. H., & Sun, D. W. (2019). Mid-infrared (MIR) spectroscopy for quality analysis of liquid foods. Food Engineering Reviews, 11, 142-158.
Su, W. H., & Sun, D. W. (2019). Mid-infrared (MIR) spectroscopy for quality analysis of liquid foods. Food engineering reviews, 11, 142-158.
Tarlak, F. (2023). The Use of Predictive Microbiology for the Prediction of the Shelf Life of Food Products. Foods, 12(24), 4461.
Totobesola, M., Delve, R., Nkundimana, J. D. A., Cini, L., Gianfelici, F., Mvumi, B., ... & Rolle, R. S. (2022). A holistic approach to food loss reduction in Africa: food loss analysis, integrated capacity development and policy implications. Food Security, 14(6), 1401-1415.
United Nations, 2021. Wasting food feeds climate change, new U.N. environment report warns. Available at: https://news.un.org/en/story/2021/03/1086402 [Accessed 06 January 2022].
Wang, Y., Yuan, Z. and Tang, Y., 2021. Enhancing food security and environmental sustainability: A critical review of food loss and waste management. Resources, Environment and Sustainability, 4, 100023.
Wold, S., Sjöström, M., & Eriksson, L. (2001). PLS-regression: a basic tool of chemometrics. Chemometrics and intelligent laboratory systems, 58(2), 109-130.
Zhang, L., Yu, Q., Zhang, M., Law, C. L., & Ma, Y. (2023). Intelligent detection of quality deterioration and adulteration of fresh meat products in the supply chain: Research progress and application. Food Bioscience, 103047.
Zhang, Y. and Haghani, A., 2015. A gradient boosting method to improve travel time prediction. Transportation Research Part C: Emerging Technologies, 58, pp.308-324.

No competing interests reported.

Download PDF

Editorial decision: Revision requested
13 Jun, 2024
Reviews received at journal
12 Jun, 2024
Reviews received at journal
06 Jun, 2024
Reviewers agreed at journal
05 Jun, 2024
Reviewers agreed at journal
03 Jun, 2024
Reviewers agreed at journal
28 May, 2024
Reviewers agreed at journal
28 May, 2024
Reviewers invited by journal
28 May, 2024
Editor assigned by journal
28 May, 2024
Submission checks completed at journal
27 May, 2024
First submitted to journal
26 May, 2024

You are reading this latest preprint version

Development of a predictive classification model and extraction of signature wavelengths for the identification of spoilage in chicken breast fillets during storage using Near Infrared Spectroscopy

Status:

Version 1

Abstract

Figures

1. Introduction

2. Material and methods

2.1. Data collection

2.2. Data processing and screening

2.2.1. Predictor Screening and Cluster Variable Analysis

2.2.2. Andrew Curve Analysis

3. Results and Discussion

4. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1