Evaluating the ability of rapid evaporative ionization mass spectrometry to differentiate beef palatability based on consumer preference

Abstract

Rapid Evaporative Ionization Mass Spectrometry (REIMS) is a type of ambient ionization mass spectrometry, which enables real-time evaluation of several complex traits from a single measurement. The objective of this study was to evaluate the capability of REIMS analysis of raw samples coupled with chemometrics to accurately identify and predict cooked beef palatability. REIMS analysis and consumer sensory evaluation were conducted for beef arm center roasts (n = 20), top loin steaks (n = 20), top sirloin steaks (n = 20), and 20% lipid ground beef (n = 20). These data were used to train predictive models for six classification sets representing different sensory traits. The maximum prediction accuracies achieved (from high to low): beefy flavor acceptance (86.25%), juiciness acceptance (83.75%), overall acceptance (81.25%), overall flavor acceptance (81.25%), grilled flavor acceptance (78.75%), and tenderness acceptance (75%). The current study demonstrates that REIMS analysis of raw meat has the potential to predict and classify cooked beef palatability.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13197-022-05562-6.

Introduction

Tenderness, juiciness, and flavor are the main factors influencing consumer choice and acceptance of cooked meat (Resurreccion, 2004). Beef palatability is often believed to be most dependent on tenderness (Miller et al. 2001). However, flavor is also considered a primary palatability factor and is shown to be of great importance when tenderness is acceptable (Behrends et al. 2005a, b). Cooking of beef generates compounds that contribute to the sensory characteristics giving beef flavor, and the amount and type of these compounds is variable depending on cooking parameters such as time, temperature, and humidity (Kerth and Miller 2015). The pre-slaughter and postmortem factors like animal breed, sex, age, feed, aging also contribute to flavor development of cooked meat (Calkins and Hodgen 2007; Khan et al. 2015).

Rapid Evaporative Ionization Mass Spectrometry (REIMS) is a type of ambient ionization mass spectrometry that was originally devised for the detection of tumor margins during cancer surgery. REIMS analysis can be performed on intact samples (no sample processing required) and yields a chemical fingerprint that, when coupled with chemometrics, enables real-time evaluation of several complex traits from a single measurement. REIMS has been successfully used in meat science (Ross et al. 2021), for characterization of production metrics such as aging method, aging time, and geographical origin of beef (Zhang et al. 2021a) and lamb (Zhang et al. 2021b). REIMS has also been used for prediction of beef quality attributes such as carcass type, production background, breed type, and muscle tenderness (Gredell et al. 2019). These research suggested the possibility of applying real-time metabolomic analysis in meat production and inspection system. However, there is limited studies using REIMS analysis of raw meat to predict cooked meat flavor. Therefore, the objective of this study was to evaluate the capability of REIMS analysis of raw meat samples coupled with chemometrics to accurately characterize cooked beef palatability based on consumer response.

Materials and methods

Product selection

Retail beef product samples were collected from five major cities in the US (Los Angeles, CA; Portland, OR; New York City, NY; Miami, FL; and Denver, CO). Within a city, beef cuts were sampled from six to eight major retail food chains stores. The objective of this sample selection process was to ensure beef products that were representative of what is typically available to consumers across the US. From each city, beef center chuck roasts, beef strip loin steaks, beef top sirloin (cap off) steaks, and 80/20 commodity ground beef were selected. From each of the five cities, ten packages of each product (4 different products) were purchased to equal 200 packages in total. After shipment with dry ice, the products were labeled, vacuum-packaged, and frozen at − 20 °C in an oxygen-barrier bag until REIMS analysis and consumer sensory panel evaluation. REIMS analysis was performed on a 2.5 cm × 10 cm section of roast, strip-loin, and sirloin steaks, and approximately 150 g of 80/20 ground beef.

Cooking procedures

Ground beef patties were formed with a hand-press patty maker, with each patty weighing 150 g. Top loin steaks, top sirloin steaks, and ground beef samples were cooked by pan grilling. Steaks and patties were cooked to a medium degree of doneness (70 °C) in a Le Creuset Signature Square Skillet Grill (Le Creuset, West Ashley, SC). Prior to cooking, the grill was preheated to 200 °C. Residues on skillet were removed between samples to avoid flavor influence. Internal temperatures were monitored periodically with a thermometer inserted into the geometric center of each steak or patty. Roasts were cooked to a medium degree of doneness (70 °C) in a conventional oven. Prior to cooking, the oven was preheated to 175 °C. Each roast was placed in an aluminum roasting pan with a rack to allow liquid to drip away from the product during cooking.

Consumer panel evaluation

The Colorado State University Institutional Review Board approved the procedures used in this study (IRB exemption # 059-19H). Samples were randomly assigned to a panel and flight order within a panel. Steak and roast samples were cut into 1.27-cm cubes using cutting guides. Ground beef (80/20) patties were cut into 6 wedges. Two cubes or wedges per sample were served on white styrofoam plates and served immediately after cooking to ensure samples were approximately 37 °C at the time of serving. Samples were identified with randomized three-digit codes and served in a randomly assigned order. Ninety-five consumers who consume meat at least three times per week participated in consumer sensory evaluations. These were performed over five-panel sessions with approximately 20 consumers per session. Within a panel session, each consumer evaluated eight different samples, and each samples were evaluated by three or four different consumers. A total of 200 samples (50 samples per type of production) were evaluated. Consumer demographic information for gender, age, ethnicity, household income, household, weekly beef consumption, and preferred doneness were collected (Supplemental Table 1). The ballot included questions for ratings of overall likeability, overall flavor, beefy flavor, grilled flavor, juiciness, and tenderness liking, using a 10-point hedonic scale anchored at both ends and the midpoint with 0 = extremely low, and extremely dislike and 10 = extremely high and extremely like. Consumers were also asked to evaluate six attributes (flavor overall acceptance, overall flavor acceptance, beefy flavor acceptance, grilled flavor acceptance, juiciness acceptance, and tenderness acceptance) with a “yes” or “no” answer and perceived acceptance levels were left to the consumers’ interpretation. Each consumer evaluated 2 ground beef patties, 2 top loin steaks, 2 sirloin steaks, and 2 chuck roast samples.

Sensory data binary classification

Within each of six sensory attributes, average liking score and acceptance responses (“yes” or “no”) for each sample were summarized. The attributes were further classified into two groups (acceptable or unacceptable) using the average liking scores of each attribute (analyzed using an analysis of variance). To be classified as “acceptable,” one of the following requirements had to be met: (1) average of liking score of a class (number of “no” in response) had to be greater than the overall mean of all classes, or (2) the upper 95% confidence interval of the mean of liking score of a class had to encompass the overall mean of all classes. Six model sets (Table 1) were defined by the following sensory attributes: overall acceptance, overall flavor acceptance, beefy flavor acceptance, grilled flavor acceptance, juiciness acceptance, and tenderness acceptance. Finally, for all six model sets, class 0 (number of “no” in response) was categorized as “acceptable,” and classes 1, 2, 3, and 4 were categorized as “unacceptable.”

Table 1.

Summary of classification groupings and number of observations used for each of the six model sets

Classifications (number of observations)	Model sets
	Overall acceptance	Overall flavor acceptance	Beefy flavor acceptance	Grilled flavor acceptance	Juiciness acceptance	Tenderness acceptance
Acceptable	53 (66.2%)	51 (63.7%)	47 (58.7%)	44 (55%)	37 (46.3%)	46 (57.5%)
Unacceptable	27 (33.8%)	29 (36.3%)	33 (41.3%)	36 (45%)	43 (53.7%)	34 (42.5%)

Rapid evaporative ionization mass spectrometry (REIMS)

Twenty samples were randomly selected from each product type (80 samples in total) and used for REIMS analysis. Prior to analysis, samples were thawed at 0–4 °C for 16–24 h. Samples (0–4 °C) were analyzed using a Synapt G2 Si Q-ToF mass spectrometer (Waters Corporation, Milford, MA), fitted with a REIMS ionization source coupled with a monopolar electrosurgical handpiece (“iKnife,” Waters Corporation) powered by an Erbotom ICC 300 electrosurgical generator (Erbe Elektromedizin GmbH, Tubingen, Germany) using “dry cut” mode at a power of 35 V. A continual flow (200 μL/min) of 200 ng/mL leucine-enkephalin was introduced directly to the REIMS source during sampling. The heater bias was set to 80 V. At least two “burns” were collected for each sample within a 2.54 × 2.54 cm square from the center of the sample, with each burn lasting approximately 3 s. Spectra were collected in positive mode ionization from 50 to 1200 m/z. Preprocessing was performed using the Abstract Model Builder (AMX) software (Beta version 1.0.2184.0, Waters Corporation), including lock mass correction (leucine-enkephalin), background subtraction using standard Masslynx preprocessing algorithms, and normalization to total ion current. Peak binning was performed at intervals of 0.5 m/z resulting in a total of 2301 bins. The bins from the collected burns were averaged to create a single value for each sample. Mass bins in the range of 550–600 m/z were excluded from the data matrix to remove the internal standard signal (leucine-enkephalin, 556.632 m/z), resulting in a final data matrix containing 2201 variables (m/z bins) and 80 observations (samples).

REIMS data analysis

Data reduction, machine learning, and evaluation of predictive models was performed within the R statistical environment (R Core Team, 2021). Data were grouped together to create the desired classifications for each model set (Table 1) defined as: overall acceptance, overall flavor acceptance, beefy flavor acceptance, grilled flavor acceptance, juiciness acceptance, and tenderness acceptance.

Data pre-processing with dimension reduction

Dimension reduction was performed using (i) feature selection (FS) or (ii) principal component analysis followed by feature selection (PCA-FS) as described by Gredell et al. (2019). PCA dimension reduction was performed using the PCA function in the package FactoMineR with unit variance scaling (Husson et al. 2017). FS was performed separately for each model set in the study (i.e., overall acceptance) using the caret R package (Kuhn 2008; Kuhn and Johnson 2013). All 2201 m/z bins were pre-processed by removing highly correlated m/z bins (Pearson’s |r|> 0.90) followed by the rfe function and finally assessed with fivefold cross validation. PCA-FS consisted of performing a similar feature selection process on the principal components rather than the 2201 mass bins with tenfold cross validation. For all model assessments performed in this study, leave-one-out cross validation refers to the removal of one sample as a validation set where the remaining samples are used as the training set. This procedure was repeated for every sample, and average prediction accuracy is recorded.

Machine learning algorithms to predict beef sensory evaluation

In total, the accuracy of fifteen machine learning algorithms was compared for each model set. These included: (1) support vector machine with a linear kernel (SVM-Linear), (2) support vector machine with a radial kernel (SVM-Radial), (3) support vector machine with a polynomial kernel (SVM-Poly), (4) random forest (RF), (5) K-nearest neighbor (Knn), (6) linear discriminant analysis (LDA), (7) penalized discriminant analysis (PDA), (8) extreme gradient boosting (XGBoost), (9) logistic boosting (LogitBoost), (10) partial least squares discriminant analysis (PLSDA), (11) stochastic gradient boosting (GBM), (12) elastic-net regularized generalized linear model (GLMNET), (13) multivariate adaptive regression spline (Earth), (14) classification and regression trees with rpart (Rpart), and (15) bagged classification and regression trees (Treebag). An initial screening of all the machine learning algorithms except PLSDA was performed using the train function from the caret package. PLSDA is not supported in the train function, and thus PLSDA models were constructed using the plsDA function (Pérez-Enciso and Tenenhaus 2003) built into the DiscriMiner package. For each of the six model sets (overall acceptance, overall flavor acceptance, beefy flavor acceptance, grilled flavor acceptance, juiciness acceptance, and tenderness acceptance), the fifteen machine learning algorithms were applied to data following the two pre-processing options, FS and PCA-FS reduction.

Leave-one-out cross validation was used to evaluate the prediction accuracy (correct predictions/total predictions) of all fifteen machine learning algorithms in order to reduce the bias and increase repeatability (James et al. 2021). The best performing model (in terms of prediction accuracy based on leave-one-out cross validation) for each model set were further optimized via parameter tuning (Table 2). Following optimization, the final prediction accuracy was recorded according to leave-one-out cross validation for each model set.

Table 2.

Summary of final prediction accuracies based on leave-one-out cross validation for the top machine learning algorithm and data reduction approach combination for each model set after parameter tuning

Model Set	Dimension Reduction approach	Number of Predictors	Machine Learning algorithm	Final Accuracy rate
Overall acceptance	PCA-FS	15 PCs	XGBoost	0.8125
Overall flavor acceptance	PCA-FS	7 PCs	XGBoost	0.8125
Beefy flavor acceptance	PCA-FS	12 PCs	XGBoost	0.8625
Grilled flavor acceptance	PCA-FS	26 PCs	Svm Poly	0.7875
Juiciness acceptance	FS	7 mass-bins	XGBoost	0.8375
Tenderness acceptance	FS	43 mass-bins	XGBoost	0.75

PCA-FS, principal component analysis followed by feature selection; FS, feature selection; PCs, principal components

Results and discussion

In the current study, four different types of beef products were used to evaluate the ability of REIMS to differentiate beef flavor acceptance (regardless of product type) based on consumer preference. Combinations of two methods of dimension reduction and fifteen machine learning algorithms were used to explore the prediction accuracy of REIMS data on six model sets (Table 1; overall acceptance, overall flavor acceptance, beefy flavor acceptance, grilled flavor acceptance, juiciness acceptance, and tenderness acceptance). The performance of each machine learning algorithm and data reduction combination was assessed in the initial screening step (Supplementary Figs. 1–6). Performance was evaluated in terms of prediction accuracy using a leave-one-out cross validation. The best performing machine learning algorithm and data reduction combinations for each model set are summarized in Fig. 1. The final prediction accuracies based on leave-one-out cross validation for the top machine learning algorithm and data reduction approach combination for each model set after parameter tuning optimization are presented in Table 2. From high to low, the maximum prediction accuracies achieved for the six classification model sets were: beefy flavor acceptance (86.25%), juiciness acceptance (83.75%), overall acceptance (81.25%), overall flavor acceptance (81.25%), grilled flavor acceptance (78.75%), and tenderness acceptance (75%).

Fig. 1.

Prediction accuracies (based on leave-one-out cross validation) for the top performing machine learning algorithm and data reduction approach combinations for each model set (overall acceptance, overall flavor acceptance, beefy flavor acceptance, grilled flavor acceptance, juiciness acceptance, tenderness acceptance). PCA-FS: principal component analysis followed by feature selection; FS: feature selection; PCs: principal components

The highest prediction accuracy (86.25%) was achieved for beefy flavor acceptance using the extreme gradient boosting (XGBoost) model with 12 principal components, followed by a 83.75% prediction accuracy for juiciness acceptance (7 mass-bins) also using the XGBoost model. Beefy flavor normally refers to the flavor associated with cooked beef or the basic meaty flavor of unseasoned beef broth, which is a major flavor attribute in beef (Adhikari et al. 2011) and generally preferred by consumers (Maughan et al. 2012; O’Quinn et al. 2016). Juiciness refers to the sensation caused by meats with higher levels of juices, which is also preferred in consumer evaluation (Maughan et al. 2012). An 81.25% prediction accuracy was achieved with XGBoost model for both overall acceptance (7 principal components) and overall flavor acceptance (12 principal components). Previous studies have demonstrated a high correlation of overall acceptance with flavor acceptance (r = 0.96; Corbin et al. 2015), and concluded that flavor had a slightly stronger impact on overall acceptance than juiciness and tenderness (Kerth and Miller 2015).

Characterizations of grilled flavor acceptance and tenderness acceptance were also evaluated, and the highest prediction accuracies (78.75% for grilled flavor acceptance and 75% for tenderness acceptance) were achieved using support vector machine with a polynomial kernel (SVM-Poly) model (26 principal components) and XGBoost model (43 mass-bins). Grilled flavor is interpreted as the flavor associated with grilled beef, which is also a major flavor attribute in beef (Adhikari et al. 2011) and is generally preferred by consumers (Maughan et al. 2012; O’Quinn et al. 2016). The tenderness acceptance prediction accuracy, while notable, is less than that previously reported by our group (90.5% accuracy) for the classification of beef tenderness based on slice shear force measurements from REIMS analysis of beef strip loin sections (Gredell et al. 2019). However, the current study represents analysis of 4 types of beef products (beef arm center roasts, top loin steaks, top sirloin steaks, and 20% lipid ground beef) which have more complex muscle composition and diverse texture than the beef strip loin evaluated in the previous study. Also, in the current study, the predictive models were trained using sensory based classification as opposed to slice shear force measurements. Therefore, the difference in prediction accuracy for tenderness acceptance classification in the current study could be attributed to the variance in tenderness resulting from product type as well as the use of a subjective evaluation for the classification grouping. A larger sample size with more observations per sample or a trained taste panel could increase the ability to discriminate palatability among samples (Ares and Varela 2017).

In addition, the model predictors are different among different prediction model sets, which assures the independence among prediction models and highlights the chance of using different model sets simultaneously to lower the misprediction probability. Furthermore, the prediction accuracy in the current study was achieved by leave-one-out cross-validation. Therefore, future studies validating the current model sets on an independent dataset should be conducted to verify the prediction rates reported in the present study.

The current study's top prediction models are built by XGBoost (overall acceptance, overall flavor acceptance, beefy flavor acceptance, juiciness acceptance, and tenderness acceptance) and SVM-Poly (grilled flavor acceptance). In agreement with the current study, our previous study (Gredell et al. 2019) revealed that the optimal machine learning algorithm assessed by predictive accuracy differed depending on the testing classification, suggesting that a “one size fits all” approach to developing predictive models from REIMS data is not appropriate. However, if additional studies are undertaken, it will be possible and it will be beneficial to have a list of recommended algorithms for the meat palatability prediction model using REIMS data.

Conclusion

The current study demonstrates, for the first time, that REIMS analysis of raw meat coupled with optimized chemometrics can characterize consumers’ overall acceptance, flavor acceptance (overall flavor and beefy flavor), and juiciness acceptance of cooked meat with greater than 81% accuracy. In other words, the probability of misprediction for each of these model sets is less than 19%, which indicates that using these four model sets simultaneously could further lower the misprediction probability and, therefore, rapidly and accurately predict the palatability of the raw beef product. Overall, this study represents a significant step in linking high throughput chemical profiling with consumer acceptance of meat products. Further REIMS analysis with larger sample size, additional products, and locations, along with an independent validation dataset, should also be conducted in the future to increase and validate the ability of REIMS to differentiate meat palatability.

Supplementary Information

Below is the link to the electronic supplementary material.

Authors’ contributions

Conceptualization, MN; Methodology, JEP; Formal Analysis, CZ, MJH-S, CLG, JEP; Investigation, CB, JL, RM, BS; Resources, MNN; Writing – Original Draft Preparation, CZ; Writing – Review & Editing, MNN, JEP; Visualization, CZ; Supervision, JEP, RD, and MNN; Project Administration, MNN; Funding Acquisition, MNN.

Funding

This work was funded by the Beef Check Off.

Availability of data and materials

The datasets generated and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Code availability

N/A.

Declarations

Conflict of interest

Authors declare that there is no conflict of interest.

Ethics approval

This project was approved by Institutional Review Board at Colorado State University, Approval number—059-19H.

Consent to participate

N/A.

Consent for publication

N/A.