Repository logo
 

Ability, repeatability, and reproducibility of rapid evaporative ionization mass spectrometry to predict beef quality attributes

Date

2022

Authors

Hernandez Sintharakao, Michael J., author
Nair, Mahesh N., advisor
Prenni, Jessica E., advisor
Morgan, James B., committee member
Sharp, Julia L., committee member

Journal Title

Journal ISSN

Volume Title

Abstract

Tenderness, juiciness, and flavor are beef quality attributes that influence consumer satisfaction eating beef. Rapid evaporative ionization mass spectrometry (REIMS) is a novel technique that provides chemical information of biological tissues with the potential to predict beef quality attributes. Two studies were conducted to evaluate the ability of REIMS to predict quality attributes of beef (study I) and to evaluate the repeatability and reproducibility of REIMS in a beef matrix (study II). In study I, USDA Select or upper two-thirds Choice (n = 42, N=84) striploins and tenderloins were collected approximately 36h post-mortem from a commercial beef abattoir. Slivers of the longissimus dorsi muscle between the 12-13th rib were collected during grading (GR, 36h post-mortem) and analyzed using REIMS. Striploins (LM) and tenderloins (PM) were cut into portions and assigned to 6 aging periods (3, 14, 28, 42, 56, and 70 days). However, only samples aged 3, 14, and 28 days were used to represent industry practices in study I. After aging, portions were cut into 2.54-cm steaks to analyze juiciness, tenderness, and 10 flavor attributes with a trained sensory panel. In addition, tenderness measures were performed using slice shear force (SSF) and Warner-Bratzler shear force (WBF). Samples were categorized by SSF, WBF, and sensory panel tenderness (PT) into "tough" and "tender"; by juiciness into "dry" and "juicy"; and by flavor into "acceptable" and "unacceptable" classes using a composite score of all flavor descriptors. Combinations of three dimensionality reduction methods (principal component analysis [PCA], feature selection, [FS], and a combination of both [PCA-FS]) with 13 machine learning algorithms were used to create classification models based on REIMS data for tenderness, juiciness, and flavor classes at the three aging periods. The predictive ability of the models was assessed with the overall accuracy resulting from 10-fold cross-validation. Among all machine learning algorithms evaluated, the maximum classification accuracies for days 3, 14, and 28 were 94, 87, and 83% for PT; 86, 85, 92% for SSF; 87, 82, and 95 for WBF; 85, 84, and 86% for juiciness; and 87, 89, and 81% for flavor classes, respectively. FS performed the best as a dimensionality reduction method for all PT, juiciness, flavor, and SSF on day 3 and WBF on days 3 and 14. PCA-FS was the best dimensionality reduction method for SSF on days 14 and 28, and WBF on day 28. Extreme gradient boosting machine learning algorithm was the highest performing algorithm for all juiciness models, flavor model on day 28, PT on days 3 and 14, SSF on days 14 and 28, and WBF on days 3. Partial least squared discriminant analysis (PLSDA) performed better for PT day 28 and flavor day 14, while elastic-net regularized generalized linear model, random forest, and support vector machine were the highest performing algorithms for SSF day 3, and WBF days 14 and 28, respectively. Results demonstrated that the chemical fingerprints obtained with REIMS could potentially be used as in situ and real-time technique to sort carcasses by flavor, juiciness, and tenderness. However, overlaps between classes affected REIMS results, and unbalanced data negatively affected model accuracies. Therefore, exploring the full potential of REIMS will require increasing the sample size and developing a sampling method that allows increased separation between sensory evaluations. Study II was performed with REIMS data from all LM and PM samples from the six aging periods (n=1008), two sets of GR samples (n=168, N=84), and quality control (QC) samples (n=29) made from homogenized ground beef. Except for the second set of GR samples, REIMS analysis of all samples was performed at Colorado State University (CSU) using a meat probe as the sampling device. Analysis of all samples was performed over 5 days, including two batches per day. GR samples were evaluated on the first day, and LM and PM data were randomly analyzed on the remaining days. QC samples were analyzed at the beginning, middle, and end of each batch. The second set of GR samples was analyzed at Texas Tech University (TTU) using different mass spectrometry (MS) instruments, technicians, and an iKnife as the sampling device. The stability of REIMS data between burns, batches, and days was evaluated with QC data. Day effect and robustness of REIMS data were evaluated with data from LM and PM samples, and interlab reproducibility was evaluated with data from GR samples. Multiple classification models of muscle type and aging were built with LM and PM data to evaluate the robustness of REIMS and day-to-day variability. Models to predict sensory attributes of beef were used to assess the robustness of REIMS with respect to interlab variability. Coefficients of variation (CV) between burns of the mass bins representing 90% of the total ion current were between 0.7 to 0.98, while the most relevant mass bins showed CV less than 0.3. Variances between batches and collection days were not significant (P < 0.05). PCA of LM and PM showed that data variability by collection day was stronger than muscle type and aging time variability. However, data could classify samples into muscle types and two distant aging times with accuracies higher than 95.6% and 91.0%, respectively. PCA of GR samples showed that data collected in both labs differed, and the predictive models developed with the CSU data did not appropriately predict the quality classes with the TTU data. REIMS collected with the meat probe provides a chemometric profile of beef samples with good repeatability and interday reproducibility but low interlab reproducibility. Consequently, optimization and standardization of sampling methods will be required to improve the interlab reproducibility of REIMS.

Description

Includes bibliographical references.
2022 Fall.

Rights Access

Subject

meat flavor
meat tenderness
machine learning
rapid evaporative ionization mass spectrometry
meat juiciness

Citation

Associated Publications