Random Forest Approach Applied to Italian-French Beef Production Systems: Sex Differences and Key Meat Standards Australia Traits Affecting Beef Eating Quality

Animal selection and data collection

Data were collected in a commercial slaughterhouse of the Associazione Zootecnica Veneta (AZoVe, Cittadella, Italy) by certified graders for the EUROP grid on one hand and the MSA carcass grading scheme (under the auspices of the International Meat Research 3G Foundation) on the other hand. The study involved 108 Charolais cattle, 39 entire males and 69 females. The average slaughter age and body weight were 518 ± 60 d and 566 ± 15 kg for females, and 510 ± 53 d and 723 ± 41 kg for males. Animals were fattened for 6 months, fed a diet rich in concentrates (Table 1), and reared in 27 Italian specialized fattening units, covering the period between May 2021 and July 2022. Although carcass grading using the MSA protocols is conducted at the 10^th ribbing site, European commercial practices typically allow grading at the 5^th ribbing site for logistical and commercial reasons (Santinello et al., 2024a). Grading at the 10^th ribbing site depreciates the carcass value, limiting data access at this location and contributing to the uneven distribution of male and female carcasses in our dataset. Indeed, in the slaughterhouse where the present study took place, the hindquarters of heifers are more frequently cut at the 10^th ribbing site to sell the loins to local butchers, whereas the hindquarters of entire males are usually sold whole to the large-scale retail sector, where they are subsequently processed. The 14-months data collection period in 27 Italian specialized fattening units considered in the present study allowed the capture a broad range of seasonal and environmental variability, improving the representativeness and usefulness of our findings.

The following MSA traits were assessed using the Australian beef chiller assessment system standards and Australian Meat standards (AUS-MEAT, 2018) in the chiller room from the surface of the Longissimus thoracis et lumborum at the 10^th thoracic vertebrae of Achilles hung carcasses, cut 24 h after slaughtering (6–7°C): 1) rib fat depth (mm), which measures subcutaneous fat; 2) total rib fat depth (mm), which includes both subcutaneous and intermuscular fat, measured manually with a ruler along the rib eye muscle at the 10^th thoracic vertebrae; 3) eye muscle area (EMA cm²), which measures the surface of Longissimus thoracis et lumborum at the 10^th thoracic vertebrae, determined using the AUS-MEAT EMA standard grid by counting the number of square centimeters of Longissimus thoracis et lumborum at the quartering site (AUS-MEAT, 2005); 4) intramuscular fat, evaluated by comparing the amount, size, and distribution of marbling with MSA reference standards, on a scale from 100 to 1190 with 10-point increments (MLA, 2006); and 5) ultimate pH (measured using a pH meter, RTD Thermometer, Delta OHM, HD2105.1, Italy). Additionally, for each carcass, the ossification score has been assessed as the degree of calcification in the sacral, lumbar, and thoracic vertebrae (Romans et al., 1994), and hump height has been measured as the distance (cm) from the highest point of the hump to the dorsal edge of the Ligamentum nuchae.

Prediction of MQ4 scores and MSA index

The EUROP-accredited technician recorded animal details (sex, birth, arrival date at the Italian fattening unit, and slaughter date) and performances (hot carcass weight, EUROP conformation, and fatness scores; EU 2013/1308, 2013), from which the length of the fattening cycle (calculated as the difference between the slaughter date and the date of arrival to the fattening unit, in days), slaughter month, and age at slaughter (days) were derived. The EUROP conformation and fatness scores were converted to a continuous 15-point scale following the method proposed by Hickey et al. (2007). Traits collected at the abattoir were used to predict the MQ4 scores for 4 commercial cuts: CUB045 (Longissimus thoracis et lumborum), STA045 (Longissimus thoracis et lumborum, anterior striploin piece), RMP131 (Gluteus medius), and TFL052 (Obliquus internus abdominis). The predictions were performed for the different commercial cuts of the same animals assuming a carcass aging time of 10 d and using the most common cooking methods for each cut as described in Santinello et al. (2024a; 2024c): grilling for CUB045 and STA045 (CUB045 GRL and STA045 GRL, respectively), stir-frying for TFL052 (TFL052 SA), and grilling and roasting methods for RMP131 (RMP131 GRL and RMP131 RST, respectively). The carcass aging time was set to 10 d, as it is the most used period at the slaughterhouse where the trial was conducted, and to maintain consistency with our previous study (Santinello et al., 2024a), since the Longissimus dorsi reaches 80% of its aging in 10 d (Pogorzelski et al., 2022).

The MSA index was calculated for a standard aging time of 5 d by summing the predicted MQ4 scores at 5 d of aging of all MSA cuts using the SP2009 version of the MSA model (McGilchrist et al., 2019). The SP2009 model includes animal sex, carcass weight (kg), hanging method, hump height (mm), ossification score, MSA marbling score, rib fat depth (mm), and ultimate pH. The MSA index was developed to provide producers with feedback on the potential meat-eating quality of their beef carcasses (McGilchrist et al., 2019). Detailed information on the MSA index and the SP2009 model is available in Polkinghorne et al. (2008), and McGilchrist et al. (2019).

Statistical analysis

Slaughter age, length of the fattening cycle, EUROP conformation and fatness scores, hot carcass weight, MSA marbling score, ossification score, eye muscle area, hump height, rib fat depth, total rib fat depth, ultimate pH, MQ4 scores, and MSA index were used as dependent variables into the following linear mixed model in SAS software v.9.4 (SAS Institute Inc., Cary, NC):

A correlation and a principal component analysis (PCA) were performed on raw data with R software (R version 2023.03.0; RStudio PBC, Boston, MA) to investigate the relationships between the investigated traits and to discriminate animals based on collected variables. Principal component analysis simplifies data interpretation by transforming correlated variables into uncorrelated principal components, effectively reducing dimensionality while retaining key sources of variability (Jolliffe and Cadima, 2016). This orthogonal transformation reduces redundancy and minimizes noise associated with overlapping variance, making it particularly useful in datasets with correlated variables, such as in the present study (Destefanis et al., 2000). Additionally, PCA can help cluster both variables and individuals, enhancing the ability to differentiate groups within complex datasets using features of several variables (Vichi and Saporta, 2009). After running the PCA, the first 2 principal components were plotted to discriminate animals according to carcass and MSA traits. Correlation analysis was conducted on the entire dataset (raw data), including both males and females. Additionally, partial-correlation analysis was performed to identify any differences or changes in correlations between the 2 groups.

Effects of sex on performance and MSA traits

Table 2 reports the least-squares means of animal and carcass characteristics (including observed MSA traits and predicted MQ4 scores) according to the sex of the animal. Males outperformed females for EUROP conformation score and hot carcass weight (P < .05; Santinello et al., 2024b), whereas no differences were reported for slaughter age and length of the fattening cycle (P > .05). Males had higher ossification scores, eye muscle area, and hump height, and females had higher EUROP fatness scores, marbling scores, rib fat depth, and total rib fat depth (P < .05), in agreement with Venkata Reddy et al. (2015). Males and females had similar ultimate pH (P < .05). The MQ4 quality scores and MSA index were higher for females than males (P < .05). The higher MSA marbling score, rib fat depth, and total rib fat depth of females compared with males are reflected also in the higher MQ4 scores for females as previously observed in different studies. For instance, the previous study of Santinello et al. (2024a) conducted in Italy, reported higher MSA marbling scores for females than males (+40 marbling points). These sex differences may be linked to the distinct hormonal profile between males and females (Schumacher et al., 2022), which favors higher muscle growth in males and higher intramuscular fat deposition in females. The hormonal status of the animals significantly impacts hot carcass weight, hump height, marbling score, and ossification score (Packer et al., 2021). Among other factors, those differences are due to testosterone production, which promotes muscle growth without a concomitant increased fat deposition in males (Lee et al., 1990).

PCA and correlation analysis

The results of the PCA analysis are depicted in Figure 1 and confirm the results of the analysis of variance. The first 2 principal components explained approximately 60% of the variability between males and females. The first 2 principal components are related to the contrast between muscle development, which is much higher in males, and fat deposition which is a characteristic of females. Indeed, females had higher EUROP fatness scores, total rib fat depth, rib fat depth, MSA marbling scores, MSA index, and MQ4 scores. Conversely, males had higher hump height, ossification scores, eye muscle area, and EUROP conformation scores. The results suggested that beef from males may have lower eating quality due to the higher hump height, ossification score, eye muscle area, and EUROP conformation score than females.

Figure 1.

Principal component analysis using animal characteristics, MSA traits, predicted meat quality scores (MQ4),¹ and MSA index.²

¹Predicted MQ4 for 5 “cut × cooking method” combinations. The model assumes that all the carcasses were Achilles hung and all cuts were assumed to be aged for 10 d and cooked according to the most common cooking method for each cut: CUB045 GRL = MQ4 score of CUB045 (Longissimus thoracis et lumborum) assuming the grilling cooking method; RMP131 GRL = MQ4 score of RMP131 (Gluteus medius) assuming the grilling cooking method; RMP131 RST = MQ4 score of RMP131 (Gluteus medius) assuming the roasting cooking method; STA045 GRL = MQ4 score of STA045 (Longissimus thoracis et lumborum, anterior striploin piece) assuming the grilling cooking method; TFL052 SF = MQ4 score of TFL052 (Obliquus internus abdominis) assuming stir-frying cooking method.

²MSA index = carcass predicted MSA score calculated as the weighed sum of the predicted MQ4 scores of all MSA cuts. The model assumes that all the carcasses were Achilles hung and all cuts were assumed to be aged for 5 d and cooked according to the most common cooking method for each cut.

Coefficients of correlation (r) between the variables estimated on the whole dataset are presented in Table 3. Slaughter age and ossification score were moderately correlated (r = .42) and were slightly positively associated with MSA marbling (r = .19 and .16 respectively), suggesting that older and mature animals tend to potentially deposit more marbling in their meat (Soulat et al., 2023). However, the ossification score was negatively correlated with MQ4 for RMP131 GRL (r = −.50) and RMP131 RST (r = −.49), and with MSA index (r = −.32), which implies that mature animals tend also to have lower meat-eating quality likely due to maturity-related collagen crosslinking formation. The EUROP conformation score was negatively correlated with fat and meat quality measurements such as EUROP fatness score (r = −.86), MSA marbling score (r = −.32), rib fat depth (r = −.26), total rib fat depth (r = −.21), and all MQ4 scores and MSA index (r = −.19 to −.52), and positively correlated with hot carcass weight (r = .82), eye muscle area (r = .72), and hump height (r = .51). On the other hand, EUROP fatness score was negatively correlated with growth-related traits such as hot carcass weight (r = −.89), eye muscle area (r = −.70), and hump height (r = −.57), and positively correlated with MSA marbling score (r = .44), total and rib fat depth (r = .19 and .28), and MQ4 scores and MSA index (r = .29 to .63). However, it is important to note that 99% of the females were classified in EUROP fatness class 3 and 97% of the males in class 2. This classification may have influenced the results.

Tables 4 and 5 report the partial correlation between EUROP fatness score and hot carcass weight, which is −.10 for females and −.51 for males. The differences in the magnitude of correlations suggest how males and females have different patterns of fat deposition. The European fatness score may be a moderate proxy for marbling level and consumer satisfaction, even if it assesses the presence of superficial fat on the carcass and not marbling in the present study. Our results partially agree with those of Bonny et al. (2016c) who reported that 3 muscles out of 17 had a positive association between EUROP fat score and tenderness. Indurain et al. (2009) assessed a correlation of .49 between the marbling score measured through image analysis and the EUROP fatness score in Pirenaica yearling bulls. In our study, MSA marbling score was positively and moderately to weakly correlated with rib fat depth and total rib fat depth (r = .33 and .27, respectively, Table 3), and with all MQ4 scores and MSA index (r = .51 to .80, Table 3). The MSA marbling was negatively correlated with eye muscle area (r = −.30), due to the dilution effects of marbling flecks into a greater surface. Hot carcass weight was negatively associated with MSA marbling (r = −.33), total and rib fat depth (r = −.14 and −.15), and all MQ4 scores and MSA index (r = −.22 to −.53), and positively associated with eye muscle area (r = .77) and hump height (r = .59). In turn, eye muscle area and hump height were negatively associated with total and rib fat depth (r = −.11 to −.30), all MQ4 scores and MSA index (r = −.18 to −.60), and positively correlated to each other (r = .38). The ossification score was positively correlated to hot carcass weight (r = .39) and eye muscle area (r = .30). Rib fat depth and total rib fat depth were positively correlated with all MQ4 scores and MSA index (r = .30 to .49). All MQ4 scores and MSA index were strongly positively correlated between them (r = .71 to .98).

Random Forest classifier applied to sex effect

The accuracy, specificity, and sensibility of the RF classifier over all iterations of 10-fold cross-validation averaged .98, .99, and .97, respectively. The average AUC of the ROC curves over all repetitions of cross-validation was .99 (Figure 2). According to MDG, the most influential variables associated with sex variability were hot carcass weight, EUROP fatness score, eye muscle area, hump height, and EUROP conformation (Figure 3). Stability of MDG values across iterations is reported in Figure 4. The RF classifier showed good performance metrics and identified, as expected, hot carcass weight, EUROP fatness score, and general muscular characteristics (eye muscle area, EUROP conformation score) as the main contributors to differences between males and females. These results confirmed the output of the PCA analysis, with males being more developed in terms of muscularity than females, and females being more prone to deposit intramuscular fat and thus to have higher beef quality (MQ4 and MSA index). The low MDG for MSA marbling and the high MDG of EUROP fatness score indicate that fat deposition in females is more concentrated in superficial fat rather than in intramuscular fat. Despite MSA marbling being significantly higher in females (Table 2), the potential for intramuscular fat deposition is not fully expressed due to the young age at slaughter. Marbling tends to develop more with increased slaughter age (Park et al., 2018), and this potential difference is expected to be even more pronounced in animals that are allowed to reach an older age, in which maximum intramuscular fat deposition occurs. This is aligned with the lower values of MDG obtained for all MQ4 scores and MSA index compared with the MDG of muscle development characteristics. Indeed, slaughter age and length of the fattening cycle are similar (Table 2), and this confirms that there are differences in marbling deposition between males and females. Moreover, the European grading scheme considers traits that are poorly or not related to the sensory properties of beef, and this can reduce consumer satisfaction (Monteils et al., 2017). Thus, the use of criteria to better reflect beef eating qualities and the optimization according to sex could improve both product consistency and consumer enjoyment.

Figure 2.

Mean ROC curve calculated on each 10-fold cross-validation, iterated 2,000 times of RF classifier applied to sex effect.

Figure 3.

Mean decrease in Gini illustrating variable importance as estimated by the RF model for sex classification, utilizing variables from Charolais cattle. The MDG values were calculated for each 10-fold cross-validation and averaged across 2,000 iterations.

Animal maturity: OSS = ossification score; SA = slaughter age (days). Fat development: RIB = rib fat thickness (mm); TRIB = total rib fat thickness (mm); MSA = MSA marbling; FATNESS = EUROP fatness score. Fattening cycle: LFC = length of the fattening cycle (days). Meat quality: pH = ultimate pH; predicted MQ4 for 5 “cut × cooking method” combinations. The model assumes that all the carcasses were Achilles hung and all cuts were assumed to be aged for 10 d and cooked according to the most common cooking method for each cut: CUB045 GRL = MQ4 score of CUB045 (Longissimus thoracis et lumborum) assuming the grilling cooking method; RMP131 GRL = MQ4 score of RMP131 (Gluteus medius) assuming the grilling cooking method; RMP131 RST = MQ4 score of RMP131 (Gluteus medius) assuming the roasting cooking method; STA045 GRL = MQ4 score of STA045 (Longissimus thoracis et lumborum, anterior striploin piece) assuming the grilling cooking method; TFL052 SF = MQ4 score of TFL052 (Obliquus internus abdominis) assuming stir-frying cooking method; MSA index = carcass predicted MSA score calculated as the weighed sum of the predicted MQ4 scores of all MSA cuts. The model assumes that all the carcasses were Achilles hung and all cuts were assumed to be aged for 5 d and cooked according to the most common cooking method for each cut. Muscle development: EMA = eye muscle area (cm²); CONF = EUROP conformation score; HUMP = hump height (cm); HCW = hot carcass weight (kg).

Figure 4.

Mean decrease in Gini averaged for each trait along iterations estimated through RF approach applied to sex effect.

Random Forest classifier applied to MSA index classes

Figure 5 shows the ROC curve of the RF classifier over all iterations of the 10-fold cross-validation applied to the MSA index classes. The average accuracy, specificity, sensibility, and AUC were .76, .78, .73, and .85, respectively. The results presented in Figure 6 indicated that MSA marbling, EUROP fatness score, ossification score, rib fat thickness, total rib fat thickness, and slaughter age had the highest MDG values and thus explained most of the variability between the 2 MSA index classes. Stability of MDG values across iterations is reported in Figure 7. Meat Standards Australia marbling has the highest MDG, and if compared with the EUROP fatness score, it is more accurate to measure predicted beef eating quality. Indeed, a lot of quality traits of meat are significantly influenced by intramuscular fat content (Hocquette et al., 2010). Previous studies have highlighted the weak and inconsistent correlation between EUROP fatness scores and meat quality attributes like marbling. The moderate correlations reported by Indurain et al. (2009) and Conroy et al. (2009) imply that although EUROP scores are informative, they may not fully capture meat quality nuances, indicating the potential value of integrating measures like the MSA marbling score, which is more closely tied to beef eating quality (Hocquette et al., 2011). Thus, probably a modern grading scheme should maximize both sensory and yield outcomes.

Figure 5.

Mean ROC curve calculated on each 10-fold cross-validation, iterated 2,000 times of RF classifier applied to MSA index classes.

Figure 6.

Mean decrease in Gini illustrating variable importance as estimated by the RF model for MSA index classification, utilizing variables from Charolais cattle. The MDG values were calculated for each 10-fold cross-validation and averaged across 2,000 iterations.

Animal maturity: OSS = ossification score; SA = slaughter age (days). Fat development: RIB = rib fat thickness (mm); TRIB = total rib fat thickness (mm); MSA = MSA marbling; FATNESS = EUROP fatness score. Fattening cycle: LFC = length of the fattening cycle (days). Meat quality: pH = ultimate pH. Muscle development: EMA = eye muscle area (cm²); CONF = EUROP conformation score; HUMP = hump height (cm); HCW = hot carcass weight (kg).

Figure 7.

Mean decrease in Gini averaged for each trait along iterations estimated through RF approach applied to MSA index classes.

Ossification score had a higher MDG value compared to slaughter age. Bonny et al. (2016a) found that ossification is a better indicator of beef eating quality than slaughter age in young European cattle. The current study showed that eye muscle area, hot carcass weight, hump height, and EUROP conformation score, having lower MDG values, reflect muscular growth instead of beef eating quality.