Body weight (BW) is widely recognized as a fundamental indicator in cattle production, contributing to the evaluation of growth performance, productivity, health conditions, and nutritional strategies across different growth stages [1–3]. To ensure optimal productivity and sustain farm profitability, livestock producers are required to regularly monitor the status of individual animals and make timely and informed decisions [4,5]. Among these decisions, early estimation of the target BW can play a pivotal role in farm management, as it is closely linked to slaughter timing and economic outcomes [2,6,7]. Delayed slaughter beyond the target BW may result in unnecessary feed costs and ultimately impair efficient resource use [8].

In recent years, the livestock industry has shown increasing interest in leveraging historical data to enhance the current productivity and decision-making [9–11]. However, in large-scale commercial operations, BW data are often difficult to collect on a regular basis because of labor constraints, equipment costs, and potential stress to animals during handling [2,5]. Even when BW records are available, such data are frequently underutilized because of the limited accessibility to appropriate analytical tools and predictive frameworks. As a result, producers are prone to relying on visual judgment, intuition, and experience-based decisions rather than evidence-based strategies. To address these limitations, predictive tools that leverage historical data offer promising approaches for improving the efficiency of beef production [12].

BW has primarily been estimated indirectly by morphometric traits with techniques ranging from simple tools such as measuring tapes [13] to advanced technologies that extract features from devices such as two-dimensional (2D) or three-dimensional (3D) cameras, ultrasound, and infrared sensors [2,14,15]. In the case of Hanwoo cattle, several studies have proposed models to estimate BW via body size measurements [16] or 3D imaging techniques [17]. To the best of our knowledge, the application of machine learning (ML) and deep learning (DL) models to predict the final BW represents a novel approach in Hanwoo steers, as previous studies have mainly focused on estimating current BW using morphometric or imaging-based methods. Using ML and DL models with sequential feeding and BW data enables time-ahead prediction of final BW without relying on specialized equipment.

A range of supervised ML algorithms such as k-nearest neighbors (KNN), random forest (RF), and eXtreme Gradient Boosting (XGBoost) have been widely applied to predict weight-related traits in beef cattle because of their ability to model complex nonlinear relationships in structured tabular data [5,16]. While conventional ML models such as KNN, RF and XGBoost often perform well on static structured datasets by effectively learning nonlinear interactions among input variables, they do not inherently model temporal dependencies in sequential data. The long short-term memory (LSTM) network is a DL model specifically designed to handle sequential data, effectively capturing both short- and long-term temporal dependencies [18,19]. Its strengths have been demonstrated across various fields including engineering [20,21], medicine [22], and animal science [23,24]. In the context of Hanwoo production, LSTM may provide a useful framework for modeling individual growth trajectories using historical data.

Therefore, this study aims to compare the prediction performance of both the ML and DL models for forecasting the final BW of Hanwoo steers at different time points prior to slaughter. By incorporating historical monthly BW and feed nutrient intake data, we hypothesize that the LSTM model will yield superior prediction accuracy due to its temporal modeling capacity.

Figure 2 illustrates the monthly distribution of BW in Hanwoo steers from 7 to 31 months of age, classified into three growth stages. As steers aged, both average and individual differences in BW markedly increased. At 7 months of age (beginning of growing stage), the average BW was 206.3 kg, which was accompanied by a SD of 27.7 kg. It increased to 417.5±34.6 kg by 17 months with greater variation among individuals. The mean BW reached 646.3±52.4 kg at 22 months and 828.5±60.0 kg at 31 months (minimum, 662 kg; maximum, 1,010 kg). These results showed increasing BW variation across animals in later stages. In particular, the observed variation in the late-stage BW reflects high individual variability, which presents a significant challenge for accurately forecasting the final BW. Greater variation among individuals implies distinct growth trajectories, which is influenced by genetic potential, nutritional responses, and environmental interactions [29]. Such individual differences complicate accurate prediction because generalized models may fail to accurately represent animals exhibiting atypical growth patterns. A previous study has highlighted that a substantial proportion of BW variation was attributable to individual animal differences rather than solely from environmental factors [29]. This underscores the necessity of incorporating individualized growth profiles for precise forecasting [30]. Therefore, models accommodating the variability through individualized growth trajectory modeling and sequential BW data inputs can enhance predictive accuracy.

The temporal distribution of nutrient intake from 7 to 31 months of age is shown in Figure 3. The amount of CPI steadily increased from 80.2±5.9 g/day at 7 months to a peak of 172.3±16.2 g/day at 20 months, then decreased to 149.3± 16.0 g/day at 31 months. The interquartile range (IQR) also expanded from 10.7 to 33.7 g/day, indicating increasing variation in CPI. A similar pattern was observed for EEI, which peaked at 40.2±4.3 g/day at 20 months and decreased afterward. NDFI displayed an earlier peak with 400.9±33.1 g/day at 12 months and decreased to 237.3±45.3 g/day by 31 months. Notably, the IQR for NDFI widened dramatically from 25.9 to 80.8 g/day in the late fattening stage. Similarly, TDNI increased from 385.8±24.9 g/day at 7 months to 807.6± 74.8 g/day at 20 months, and then slightly declined to 728.1± 40.1 g/day at 31 months.

These findings support the interpretation that differences between individual animals become more pronounced in the late fattening stage. In Hanwoo production systems, high concentrate diets are typically given during the late fattening stage to increase intramuscular fat content in carcass [31]. However, feeding high concentrate diets during the late fattening stage in Hanwoo steers could reduce feed intake because of possible occurrence of subacute ruminal acidosis caused by rapid starch fermentation and subsequent declines in rumen pH [32,33]. The feed intake variation in the later stages found in the present study apparently resulted from intake reduction in the animals.

Table 4 presents the predictive performance of four models: KNN, XGBoost, RF, and LSTM. The models were evaluated with six different input configurations (I1–I6) representing prediction time points from 17 to 1 month(s) before slaughter. The predictive accuracy was improved as the prediction point approached slaughter in all the models. Among the models, LSTM consistently outperformed others, with R² values improving from 0.60 (I1) to 0.97 (I6) and RMSE values decreasing from 52.80 kg to 22.62 kg in the test set. The highest prediction accuracy was observed for I6 in LSTM. The RF model achieved the highest prediction accuracy at I6 among the ML models, second only to the LSTM, and showed robust generalization. XGBoost exhibited lower predictive accuracy compared with RF across all input configurations, and signs of overfitting were observed from I3 to I6. The lowest predictive accuracy was shown in KNN with its best performance reaching only R² = 0.75 and RMSE = 56.55 kg at I6, which reflected weakest generalization.

The superior performance of the LSTM model appears to stem from its ability to capture sequential dependencies and temporal dynamics, enabling effective use of extended historical sequences [24]. This capability was especially beneficial in the late fattening stage, where increased variability in BW and nutrient intake makes forecasting more difficult. In the present study, the effectiveness of complex models like LSTM was highly dependent on their ability to align with the structure of the input data, and these findings underscore the importance of both input sequence length and appropriate model selection in enhancing prediction accuracy for growth in Hanwoo steers. Forecasting the final BW in advance offers practical advantages to farm-level decision-making including slaughter timing and feeding program adjustment [12]. While the use of longer input sequences can be advantageous for achieving the highest accuracy, this presents a practical compromise for on-farm management. For example, predictions made one month before slaughter (I6) may give a limited time for management such as dietary intervention in farms. Instead, mid-range predictions of I4 (6 months) and I5 (3 months) would allow more time for farm management with relatively high prediction accuracy (R² values of 0.83 and 0.90 and RMSEs of 35.84 kg and 33.12 kg, respectively). Among ML models, the robust generalization of the RF model was indicated by its minor discrepancies between the training and test sets. In contrast, XGBoost showed a tendency toward overfitting as its performance was higher in the training set compared to the test set, and this was more apparent with longer data sequences. The overfitting occurs when a model learns the noise in the training data as well as the underlying patterns and consequently reduces the ability to generalize to new data. This indicates that boosting methods such as XGBoost may be more prone to noise and weaker generalization compared to bagging methods such as RF that provided more stable and consistent performance in the present study [34]. High variation and noise in BW and nutrient intake in our dataset may have favored the variance-reduction strategy of RF over the boosting approach of XGBoost. The overfitting is a significant issue in model development because it hinders the ability of the model to make reliable predictions on unseen data. This can lead to inaccurate management decisions under practical farm conditions. Meanwhile, the large SDs of KNN in the training meant unstable learning performance. This instability is likely to have contributed to the low predictive accuracy observed in the test set, which ultimately reflected weak generalization.

The previous studies in Hanwoo have focused on the current BW estimation with limited attention to time-ahead prediction. For instance, Jang et al [17] estimated the current BW of Hanwoo cattle using 3D imaging techniques including top-view images captured by time-of-flight and stereo vision cameras. Their multiple linear regression models that incorporated body dimensions and age yielded RMSEs between 51.4 and 62.0 kg. Dang et al [16] applied ML models to predict current BW based on ten body measurements including wither height, body length, and chest girth. They achieved higher predictive accuracy compared to the study of Jang et al [17], with RMSEs ranging from 24.75 to 28.55 kg. Despite these improvements, their reliance on body measurements or specialized 3D imaging systems presented practical limitations for real-farm adoption due to cost, labor, and infrastructure burdens [2,35]. In contrast, the present study demonstrated a more accessible and cost-effective approach by utilizing historical data collected from a farm without the need for additional imaging devices or specialized measurement equipment. This highlights the practical advantage of our method for implementation in commercial Hanwoo production systems.

The time-ahead prediction of carcass traits has been explored in studies with different breeds of beef cattle. For instance, Duwalage et al [12] predicted carcass weight in grass-fed beef cattle using seven phenotypic traits: weaning weight, weight gain since weaning, time since weaning, breed, sex, weaning season (wet or dry), and farm property. They developed predictive models using boosted regression trees (a ML method) and multiple linear regression, and reported RMSE values of 11.65, 15.34, 20.76, and 25.06 kg when predicting carcass weight at 1, 3, 9–10 months before slaughter, and at weaning, respectively. Similarly, Alonso et al [36] used zoometric measurements to forecast carcass weight and achieved mean absolute percentage error values of 3.3% and 3.9% at 1 and 3 months, respectively, prior to slaughter. These studies highlighted the feasibility of time-ahead prediction in beef production although their targets (carcass weight) and input features differed from the present study’s live BW and nutrient intake data. The values of RMSEs in the LSTM of the present study were comparable to or slightly higher than the carcass weight prediction RMSEs reported by Duwalage et al [12] and Alonso et al [36]. This is reasonable because live BW is more dynamic and potentially more variable than carcass weight due to temporary factors.

In future studies, the integration of more accurate dry matter intake records and sensor-based BW monitoring would improve both prediction accuracy and scalability [37,38]. Additionally, future research should consider explainability and user-centered design to promote adoption in field settings. This is particularly important because DL models are inherently black-box systems, making it difficult for end users to interpret their internal reasoning processes [39]. Explainable artificial intelligence methods can clarify how specific input features influence predictions, thereby improving model transparency [40]. By integrating these interpretability techniques into decision-support tools, predictive models can become not only more accurate but also more trustworthy and actionable for commercial livestock management.

We developed and compared four predictive models (KNN, XGBoost, RF, and LSTM) to predict the final BW of Hanwoo steers before slaughter using monthly BW and estimated nutrient intake data. Among the tested models, the LSTM model yielded the highest prediction accuracy across all input configurations. When excluding LSTM, RF, XGBoost, and KNN ranked in descending order of performance based on test R² and RMSE across all input configurations. Prediction accuracy improved as the forecasting time point moved closer to the target slaughter age. These models can serve as practical tools to support timely and data-driven decision-making in Hanwoo production systems. In particular, the LSTM model shows strong potential for integration into farm-level decision-support systems and precision feeding programs, which can contribute to more efficient management and optimized resource utilization in commercial Hanwoo production.