Product selection

The samples were collected from 200 sheep carcasses at 3 US Department of Agriculture-certified slaughterhouses in California and Colorado. The sampling was designed to represent 2 age groups (n = 99 lambs, n =101 yearlings), 2 diet groups (n = 101 grass, n = 99 grain), and 2 gender groups (n = 96 females, n = 104 males). Before slaughter, animals were identified, classified by age, diet, and gender, and tagged with a unique carcass identifier to ensure traceability throughout sampling and analysis. At 45 min postmortem, approximately 5 g of biceps femoris tissue, including both lean muscle and external fat, were excised from a single deep incision made on 1 side of each carcass and individually packed. Samples were immediately frozen in dry ice and shipped to Colorado State University, where they were stored at −80°C. At 24 h postmortem, 1 leg from each of these carcasses was collected and transported to Texas Tech University on ice. At 7 d postmortem, the legs were boned and trimmed of all fat (external and seam), finely ground, shaped into slider-sized patties (5.08 cm in diameter), vacuum-packed, and frozen at −80°C. The patties were shipped on dry ice to Colorado State University for analysis. For each carcass, 1 composite, comprising lean tissue and external fat, and 1 patty were allocated to I-Knife REIMS analysis. An additional patty was allocated to Meat Probe REIMS analysis. Furthermore, 6 patties were allocated for sensory analysis and were frozen (−20°C) in vacuum bags until sensory analysis.

Consumer panel evaluation

The procedures used in this study were approved by the Colorado State University Institutional Review Board (exemption #2039). A total of 200 consumers were recruited from Fort Collins, Colorado, and nearby regions for the sensory panels. Table 1 provides the demographic breakdown of the 200 participants in the sensory panels. Sensory panels were conducted at Colorado State University in groups of 20 panelists per session. Patties were thawed (2–4°C) 24 h prior to consumer panels and cooked in a combi-oven (Model SCC WE 61 E; Rational, Landsberg am Lech, Germany) set to a temperature of 204.5°C with 0% humidity. The internal temperature was monitored by probing the geometric center of each patty using a type-K thermocouple thermometer (AccuTuff 34032; Cooper-Atkins Corporation, Middlefield, CT). Patties were cooked for approximately 6 min until the internal temperature reached a minimum of 71°C and were immediately served to consumers as full patties. A napkin, plastic fork and knife, toothpick, expectorant cup, apple juice (10% apple juice dilution), water, and unsalted crackers were provided to each panelist to use as palate cleansers.

Each consumer evaluated 1 familiarization sample followed by 6 test samples (n = 7 samples/panelist) in a predetermined and balanced order, representing variation in animal age class and production background, on an electronic tablet (iPad Pro, Model A1670; Apple Inc., Cupertino, CA) using a digital survey (Qualtrics core XM; Qualtrics Software, Provo, UT). The familiarization sample was provided to help participants acclimate to the digital survey interface and the sensory characteristics of lamb prior to formal evaluation. Participants evaluated each sample for flavor intensity, flavor liking, and overall liking using a continuous line scale that was initially anchored at the midpoint. The scale ranged from 0 to 100, where for flavor intensity, 0 was “extremely low” and 100 was “extremely high,” and for flavor liking and overall liking, 0 was “extremely dislike” and 100 was “extremely like.” Panelists were also asked to assess with a binary response (“yes” or “no”) for flavor intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance, leaving acceptance levels to consumers’ interpretation.

Sensory data classification

In addition to the 3 carcass background classifications (age, diet, and gender), 5 sensory attributes were further classified according to the consumer evaluation (Table 2). Flavor intensity responses were also divided into 3 categories: (1) mild, (2) medium, and (3) intense. Flavor intensity groupings were developed based on a normal standard curve using plus or minus 1 standard deviation from the mean as separation. For the remaining attributes, the average overall liking score and binary responses (“yes” or “no”) were summarized for each sample. The samples were first placed in different classes (number of “yes” in response) within an attribute, and the average overall liking score for each class (number of “yes” in response) was used to further classify the sample classes into binary groups (“acceptable” or “unacceptable”). For each attribute, samples classified as “acceptable” met at least 1 of these 2 requirements: (1) average of overall liking score of the class (number of “yes” in response) was greater than the overall mean of all classes, or (2) the upper 95% CI of the mean of overall liking score of the class included the overall mean of all classes. Four model sets (Table 2) were defined by the following sensory attributes: flavor intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance. Finally, for intensity acceptance, flavor acceptance, and overall acceptance, classes (number of “yes” in response) 5 and 6 were classified as “acceptable,” and classes 0, 1, 2, 3, and 4 were categorized as “unacceptable.” For the presence of off-flavor, classes (number of “yes” responses) 0 and 1 were classified as “acceptable,” while classes 2, 3, 4, 5, and 6 were classified as “unacceptable.”

Rapid evaporative ionization mass spectrometry

Samples were thawed at a temperature between 0°C and 4°C for 16 h to 24 h before REIMS analysis. Chemical fingerprints of fat tissue, lean tissue, and ground patties were acquired using a REIMS ionization source fitted to a Synapt G2 Si Q-ToF mass spectrometer (Waters Corporation, Milford, MA). An “I-Knife” (Waters Corporation), a monopolar electrosurgical handpiece, or a “Meat Probe” (Waters Corporation), a bipolar electrosurgical handpiece, was used as the sampling device and powered with an Erbotom ICC 300 electrosurgical generator (Erbe Elektromedizin GmbH; Tübingen, Germany) set in “dry cut” mode. The generator was set to 30 V for lean and patty samples or to 50 V for fat samples.

A solution of 200 ng/mL leucine-enkephalin was continuously introduced to the REIMS source at 200 μL/min during sampling. The heater bias was set to 80 V. Sample collection was performed by making at least 5 “burns” within a 2.54 × 2.54 cm square on the center of the samples. All burns were performed by 1 person and lasted approximately 3 s to maintain consistency. Mass spectra from 50 to 1500 m/z were collected in negative-ion mode. Data preprocessing included lock mass correction (leucine-enkephalin), background subtraction with standard Masslynx preprocessing algorithms, and total ion current normalization using the AMX recognition software version 1.0.2184.0 (Waters Corporation). Peaks between the 50 m/z to 1500 m/z range were binned into 0.5 m/z intervals, resulting in 2900 bins per sample. For each sample, data from the 5 replicate burns were averaged across corresponding bins to generate a single representative dataset. To minimize interference from the internal standard leucine-enkephalin (m/z 554.632), all bins within the 550 m/z to 600 m/z range were excluded from the final data matrix.

To evaluate the predictive capacity of REIMS for cooked sheep meat flavor and carcass characteristics using chemical fingerprints acquired from raw samples by the I-Knife, mass-bins in the range of 100 to 550 and 660 to 1000 were selected. The selection resulted in a data matrix of 1700 variables (m/z bins) and 200 observations for each type of sample (lean, fat, and patty). For comparison of the data generated by the Meat Probe and the I-Knife, mass-bins in the range of 50 to 550 and 660 to 1500 were used for further analysis. The selection resulted in a data matrix of 2800 variables (m/z bins) and 200 observations for each meat patty. The time required for REIMS analysis per sample using both sampling devices was also recorded.

Rapid evaporative ionization mass spectrometry data analysis

Data dimensionality reduction, machine learning modeling, and predictive performance evaluation were conducted using the R statistical environment (R Core Team, 2021). Data were organized into specific classification categories for each model set, as outlined in Table 2. These classifications included age, diet, gender, flavor intensity level, flavor intensity acceptance, flavor acceptance, presence of off-flavors, and overall acceptance.

Data preprocessing with dimension reduction

Feature selection (FS) or principal component analysis followed by FS (PCA-FS) as described by Gredell et al. (2019) was used for dimension reduction. The PCA dimension reduction was performed using the PCA function in the FactoMineR package with unit variance scaling (Husson et al., 2017), and FS was performed separately for each model set in the study (i.e., overall acceptance) using the rfe function from the caret R package (Kuhn, 2008; Kuhn and Johnson, 2013). The FS was performed individually for each model by removing highly correlated mass-bins (Pearson’s |r| > 0.90) from the initial 2800 mass-bins, followed by the rfe function, and a 5-fold cross-validation. The PCA-FS was performed by applying a similar process of the FS to the principal components instead of the mass-bins and using a 10-fold cross-validation instead of the 5-fold cross-validation.

Machine learning algorithms to predict carcass background and meat sensory evaluation

A total of 15 different machine learning algorithms were evaluated for each of the 8 classification model sets (age, diet, gender, flavor intensity level, flavor intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance). The algorithms were: (1) support vector machine (SVM) with a radial kernel (SVM Radial), (2) SVM with a linear kernel (SVM Linear), (3) SVM with a polynomial kernel (SVM Poly), (4) K-nearest neighbors (Knn), (5) random forest (RF), (6) linear discriminant analysis (LDA), (7) penalized discriminant analysis, (8) logistic boosting, (9) extreme gradient boosting (XGBoost), (10) stochastic gradient boosting (GBM), (11) elastic-net regularized generalized linear model (GLMNET), (12) multivariate adaptive regression spline, (13) classification and regression trees with rpart, and (14) bagged classification and regression trees. These algorithms were also initially screened using the train function from the caret package.

For I-Knife tissue comparison (fat vs. lean vs. patty), partial least-squares discriminant analysis (PLSDA) was included as the 15th machine learning algorithm. Since PLSDA algorithm was not supported in the train function, these models were trained using the plsDA function from the DiscriMiner package (Pérez-Enciso and Tenenhaus, 2003). For electrode comparison, a generalized linear model (GLM) was included as the 15th machine learning algorithm using the train function from the caret package. Prediction accuracy of all the models was evaluated using leave-one-out cross-validation to reduce bias and have high repeatability (James et al., 2021). For each model set, the prediction accuracy using each combination of preprocessing method and machine learning algorithm was recorded. Regardless of preprocessing options, the best performance of each machine learning algorithm, based on the prediction accuracy, for each model set was summarized for tissue comparison (Figure 1) and electrode comparison (Figure 2). The best-performing model for each model set was summarized for tissue comparison (Table 3) and electrode comparison (Table 4).

Figure 1.

Prediction accuracy by tissue type (based on leave-one-out cross-validation) for the top-performing machine learning algorithm and data reduction approach combinations for age, diet, gender, and flavor intensity level classification using I-Knife. Earth, multivariate adaptive regression spline; FS, feature selection; GBM, stochastic gradient boosting; GLM, generalized linear model; GLMNET, elastic-net regularized GLM; Knn, K-nearest neighbors; LDA, linear discriminant analysis; LogitBoost, logistic boosting; PCA-FS, principal component analysis followed by feature selection; PDA, penalized discriminant analysis; RF, random forest; Rpart, classification and regression trees with rpart; SVM, support vector machine; SVM Linear, SVM with a linear kernel; SVM Poly, SVM with a polynomial kernel; SVM Radial, SVM with a radial kernel; TreeBag, bagged classification and regression trees; XGBoost, extreme gradient boosting.

Figure 2.

Prediction accuracy by tissue type (based on leave-one-out cross-validation) for the top-performing machine learning algorithm and data reduction approach combinations for flavor intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance classification using I-Knife. Earth, multivariate adaptive regression spline; FS, feature selection; GBM, stochastic gradient boosting; GLM, generalized linear model; GLMNET, elastic-net regularized GLM; Knn, K-nearest neighbors; LDA, linear discriminant analysis; LogitBoost, logistic boosting; PCA-FS, principal component analysis followed by feature selection; PDA, penalized discriminant analysis; RF, random forest; Rpart, classification and regression trees with rpart; SVM, support vector machine; SVM Linear, SVM with a linear kernel; SVM Poly, SVM with a polynomial kernel; SVM Radial, SVM with a radial kernel; TreeBag, bagged classification and regression trees; XGBoost, extreme gradient boosting.

Top-performing I-Knife rapid evaporative ionization mass spectrometry prediction models

This study used I-Knife REIMS to analyze 3 distinct tissue types to distinguish carcass characteristics and sheep meat flavor profiles based on consumer preferences. To assess the predictive performance of the REIMS data, 8 modeling scenarios were explored, covering variables such as age, diet, sex, flavor intensity, acceptance of flavor intensity, overall flavor acceptance, presence of off-flavors, and overall liking, using a combination of 2 dimensionality reduction methods and 15 different machine learning algorithms (Table 2). The performance of each machine learning algorithm and data reduction combination was assessed during the initial screening phase, with prediction accuracy evaluated using leave-one-out cross-validation.

The maximum prediction accuracy of data generated by I-Knife REIMS, based on leave-one-out cross-validation, varied across different model sets and sample types (Table 3). The maximal prediction accuracies from models based on fat tissue REIMS analysis were as follows: 76.5% for age, 94.5% for diet, 73% for gender, 73.5% for intensity level, 82% for intensity acceptances, 81% for flavor acceptance, 93% for off-flavor presence, and 81% for overall acceptance. The maximal prediction accuracies from models based on lean tissue REIMS analysis were as follows: 82% for age, 85% for diet, 71.5% for gender, 73% for intensity level, 84% for intensity acceptances, 80.5% for flavor acceptance, 75% for off-flavor presence, and 84% for overall acceptance. The maximal prediction accuracies from models based on patty REIMS were as follows: 78.5% for age, 92.5% for diet, 76% for gender, 69.5% for intensity level, 94% for intensity acceptances, 79.5% for flavor acceptance, 75.5% for off-flavor presence, and 83% for overall acceptance.

Regardless of tissue type, from high to low, the maximum prediction accuracies of the 8 classification model sets were the following: 94.5% for diet, 94% for flavor intensity acceptance, 93% for off-flavor presence, 85.1% for overall acceptance, 82% for age, 81% for flavor acceptance, 76% for gender, and 73.5% for flavor intensity level. For each classification model set, the maximum prediction accuracy achieved by I-Knife REIMS varied among different tissue types (Table 3). Fat tissue had higher maximum prediction accuracies for diet (94.5%), off-flavor presence (93%), flavor acceptance (81%), and flavor intensity level (73.5%), while lean tissue had higher maximum prediction accuracies for overall acceptance (85.1%) and age (82%). Patty had higher maximum prediction accuracies for flavor intensity acceptance (94%) and gender (76%).

In the current study, regardless of tissue type, the maximum prediction accuracies of diet, flavor intensity acceptance, off-flavor presence, overall acceptance, age, and flavor acceptance classification were greater than 80%. To the best of our knowledge, this study is the first to demonstrate that REIMS analysis of raw meat coupled with machine learning algorithms can accurately predict cooked sheep meat flavor and animal age. Previous studies have demonstrated the potential to predict meat quality grade and production background information with I-Knife REIMS data with various machine learning algorithms. For example, beef quality grade, production background, breed type, and muscle tenderness were predicted with maximal accuracies of 81.5% using LDA, 99% using SVM Linear, 85% using SVM Radial, and 90.5% using XGBoost, respectively (Gredell et al., 2019). Additionally, using coupling orthogonal projection to latent structures-discriminant analysis (OPLS-DA), aging method (straight 21-d dry aging vs. 7-d dry aging plus 14-d wet aging), and aging time (nonaged vs. 21-d aging) were classified by I-Knife REIMS with 85% and 95% accuracy in raw beef (Zhang et al., 2021a). A recent I-Knife REIMS study also classified aging methods with 95% accuracy and dehydration level with 82% accuracy in lamb meat using the OPLS-DA model (Zhang et al., 2023). For consumer preference prediction, I-Knife REIMS had more than 80% accuracy for specific flavor acceptance, overall flavor acceptance, juiciness acceptance, and overall product acceptance in various beef products (Zhai et al., 2022b). I-Knife REIMS also demonstrated high accuracy in differentiating instrumental tenderness, juiciness, and overall flavor acceptance in trained panel (Hernandez-Sintharakao et al., 2023).

Flavor prediction using I-Knife rapid evaporative ionization mass spectrometry of early postmortem samples

Muscle-to-meat conversion is an intensive biochemical process affecting the development of meat color (Zhai et al., 2022a; Zhu et al., 2025), tenderness (Melody et al., 2004; Schulte et al., 2023), and water-holding capacity (Melody et al., 2004). The metabolite profile of the meat can also shift during the postmortem period through postmortem glycolysis (Yu et al., 2019), proteolysis (Melody et al., 2004), and lipolysis (Monin et al., 2003), which can further affect the final product’s flavor performance (Young et al., 2003). Previous studies indicated that meat with differing quality attributes has different biochemical fingerprints at early postmortem and becomes increasingly divergent during the later postmortem period (Yu et al., 2019; Zhai et al., 2020). Therefore, the rapid onsite identification of postrigor meat flavor using prerigor muscle tissue is theoretically viable. In the current study, except for flavor intensity acceptance and gender, I-Knife REIMS data from fat and lean tissue collected at 45 min postmortem achieved higher maximal prediction accuracies than from meat patty (94.5% for diet, 93% for off-flavor presence, 84% for overall acceptance, 82% for age, 81% for flavor acceptance, and 73.5% for flavor intensity level). These results indicate that REIMS data collected from both lean and fat tissue at 45 min postmortem are reflective of 7 d postmortem meat sensory data and carcass background.

Taken together, these results demonstrate that REIMS is a promising technique for rapid, online prediction of lamb quality based on prerigor sampling. As REIMS does not require sample preparation, it enables real-time decision-making and efficient carcass sorting based on consumer-relevant traits, offering a practical and scalable solution for the lamb industry.

I-Knife rapid evaporative ionization mass spectrometry prediction across tissue types

For carcass background prediction, the highest prediction accuracy was achieved for diet classification by fat tissue (94.5%) using the SVM Poly (80 mass-bins) model (Table 3; Figure 1). Achievement of high prediction accuracy using fat tissue could be attributed to the significant impact of diet on the fatty acid composition of sheep meat (Luciano et al., 2009; Yang et al., 2022). The second highest prediction accuracy was achieved for age classification by lean tissue (82%) using the SVM Poly (75 mass-bins) model (Table 3; Figure 1). This might be attributed to the metabolomic profile change in sheep’s muscles as animals grow because older sheep exhibit higher levels of amino acids, reducing sugars, and free fatty acids (Wang et al., 2025). The next highest prediction accuracy was achieved for gender classification by patty (76%) using the SVM Radial model (29 principal components; Table 3; Figure 1). Previous studies have indicated that the effect of gender on muscle metabolome, proteome, and lipidome is dependent on the diet and age of livestock (Young et al., 1997; Bednárová et al., 2013; Picard et al., 2019; Zhang et al., 2022b; Wang et al., 2025). These interactive effects might have led to the relatively low prediction accuracy of sheep gender in our study compared to other attributes. Moreover, castration also has an effect on muscle/meat molecular fingerprints (Zhou et al., 2011). The current classification model for gender was binary (male or female) and, thus, did not separate rams from wethers, which might also explain the prediction performance.

For cooked sheep meat flavor prediction, the highest prediction accuracy was achieved for flavor intensity acceptance by patty (94%) using the GLM (1474 mass-bins; Table 3; Figure 2). The second highest prediction accuracy for sheep meat flavor was achieved for off-flavor presence by fat tissue (93%) using the GLM (1295 mass-bins; Table 3; Figure 2). Previous studies have indicated that the secondary lipid oxidation products are the major source of off-flavor in meat products (Domínguez et al., 2019), which could explain the highest prediction accuracy of off-flavor using fat tissue. The next highest prediction accuracy was achieved for overall acceptance by lean tissue (84%) using the SVM Poly model (35 principal components; Table 3; Figure 2). This indicates that lean sheep meat could be the most reliable tissue source to evaluate the overall sheep product acceptance, which could be attributed to the contribution of amino acids or peptides to the positive flavor attributes (Ma et al., 2020; Ding et al., 2024). Characterizations of flavor acceptance were also evaluated, and the highest prediction accuracy (81%) was achieved by fat tissue using GBM model (49 principal components; Table 3; Figure 2). The highest prediction accuracy for flavor intensity level was achieved using SVM Poly model to 73.5% by fat tissue with 23 principal components (Table 3; Figure 1). Similar to the off-flavor prediction, these higher prediction performances of fat tissue are likely related to the contribution of lipids to the cooked meat flavor (Dinh et al., 2021). Future studies are warranted to investigate the mechanism behind high prediction accuracy when using different tissue types.

Comparison between I-Knife and Meat Probe rapid evaporative ionization mass spectrometry

In this study, 2 types of REIMS electrodes were used to distinguish carcass background and sheep meat flavor according to consumer preference. To assess the predictive accuracy of REIMS data across 8 classification model sets, combinations of 2 dimensionality reduction methods and 15 machine learning algorithms were applied (Table 2; age, diet, gender, flavor intensity level, flavor intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance).

The highest prediction accuracies for each classification, based on data generated through I-Knife REIMS analysis (Table 4), were as follows: 75% for age, 92% for diet, 73% for gender, 73% for intensity level, 84.5% for intensity acceptance, 81.5% for flavor acceptance, 78% for off-flavor presence, and 84% for overall acceptance. The highest prediction accuracies for each classification, based on data generated through Meat Probe REIMS analysis, were as follows: 81.5% for age, 90% for diet, 72.5% for gender, 71% for intensity level, 86% for intensity acceptance, 80% for flavor acceptance, 77.5% for off-flavor presence, and 85% for overall acceptance. Overall, data generated using the Meat Probe resulted in models with similar or better accuracies of carcass background (age, diet, and gender) and consumer preference (intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance) compared to models based on data generated using the I-Knife.

For carcass background prediction, the highest prediction accuracy was achieved for diet classification by I-Knife (92%) and Meat Probe (90%) using the SVM Poly (34 mass-bins) model and SVM Linear (29 principal components) model, respectively (Table 4; Figure 3). The second highest prediction accuracy was achieved for age classification by I-Knife (75%) and Meat Probe (81.5%) using the Knn (1949 mass-bins) model and SVM Poly (86 principal components) model, respectively (Table 4; Figure 3). The next highest prediction accuracy was achieved for gender classification by I-Knife (73%) and Meat Probe (72.5%) using the RF (79 principal components) model and XGBoost (17 principal components) model, respectively (Table 4; Figure 3).

Figure 3.

Prediction accuracies by electrode (I-Knife vs. Meat Probe; based on leave-one-out cross-validation) for the top-performing machine learning algorithm and data reduction approach combinations for age, diet, gender, and flavor intensity level classification. Earth, multivariate adaptive regression spline; FS, feature selection; GBM, stochastic gradient boosting; GLMNET, elastic-net regularized generalized linear model; Knn, K-nearest neighbors; LDA, linear discriminant analysis; LogitBoost, logistic boosting; PCA-FS, principal component analysis followed by feature selection; PDA, penalized discriminant analysis; PLSDA, partial least-squares discriminant analysis; RF, random forest; Rpart, classification and regression trees with rpart; SVM, support vector machine; SVM Linear, SVM with a linear kernel; SVM Poly, SVM with a polynomial kernel; SVM Radial, SVM with a radial kernel; TreeBag, bagged classification and regression trees; XGBoost, extreme gradient boosting.

For cooked flavor prediction, the highest prediction accuracy was achieved for flavor intensity acceptance by I-Knife (84.5%) and Meat Probe (86%) using the XGBoost (33 principal components) model and SVM Linear (29 principal components) model, respectively (Table 4; Figure 4). The second highest prediction accuracy was achieved for overall acceptance by I-Knife (84%) and Meat Probe (85%) using the XGBoost (71 mass-bins/33 principal components) model and PLSDA (114 mass-bins) model, respectively (Table 4; Figure 4). The next highest prediction accuracy was achieved for flavor acceptance by I-Knife (81.5%) and Meat Probe (80%), both using the XGBoost model with 63 mass-bins and 49 principal components, respectively (Table 4; Figure 4). Characterizations of flavor intensity level were also evaluated, and the highest prediction accuracies (73% for I-Knife and 71% for Meat Probe) were achieved using the GBM model (76 mass-bins) and XGBoost model (995 mass-bins; Table 4; Figure 3). The highest prediction accuracy for off-flavor presence was achieved using SVM Poly model, reaching 78% accuracy with the I-Knife using 40 principal components and 77.5% with the Meat Probe using 33 principal components (Table 4; Figure 4).

Figure 4.

Prediction accuracies by electrode (I-Knife vs. Meat Probe; based on leave-one-out cross-validation) for the top-performing machine learning algorithm and data reduction approach combinations for flavor intensity acceptance, flavor acceptance, off-flavor presence, and overall acceptance classification. Earth, multivariate adaptive regression spline; FS, feature selection; GBM, stochastic gradient boosting; GLMNET, elastic-net regularized generalized linear model; Knn, K-nearest neighbors; LDA, linear discriminant analysis; LogitBoost, logistic boosting; PCA-FS, principal component analysis followed by feature selection; PDA, penalized discriminant analysis; PLSDA, ; RF, random forest; Rpart, classification and regression trees with rpart; SVM, support vector machine; SVM Linear, SVM with a linear kernel; SVM Poly, SVM with a polynomial kernel; SVM Radial, SVM with a radial kernel; TreeBag, bagged classification and regression trees; XGBoost, extreme gradient boosting.

Analysis using the Meat Probe required 45 s per sample (5 readings), whereas analysis using the I-Knife required 90 s per sample (5 readings). Also, REIMS analysis using I-Knife required the cleaning procedure around every 80 samples (400 readings), while REIMS analysis using Meat Probe required a cleaning procedure around every 160 samples (800 readings), which indicated that Meat Probe generated a cleaner signal than I-Knife.