Evaluation of Machine Learning Models for TT-POCUS Based Prognostic Prediction of Did-Not-Finish and First-Treatment Failure Cattle Diagnosed with Bovine Respiratory Disease

Simple Summary

Bovine respiratory disease (BRD) is a major cause of negative health outcomes in feedyard cattle. This syndrome results in significant economic and animal welfare concerns. Better information on BRD prognosis, or expected outcome following treatment, at the time of individual animal treatment could improve management decisions of diseased cattle. Targeted thoracic ultrasound is a modality to measure several variables associated with BRD and potential post-treatment outcomes. The objective of this study was to assess the accuracy of machine learning tools to predict post-treatment outcomes using data collected from targeted thoracic ultrasound. A total of 1422 first-BRD-treatment cattle and 150 head with multiple treatments were enrolled. Six machine learning algorithms were trained using ultrasound data. Algorithm performance varied by dataset, but overall, using variables from thoracic ultrasound provided meaningful prognostic information.

Abstract

Bovine respiratory disease remains the leading cause of morbidity and mortality in feedyard cattle and is associated with substantial economic and animal welfare consequences. Prognostic tools remain limited, but combining objective diagnostic measures with predictive modeling may improve clinical decision-making. This study evaluated the performance of machine learning models incorporating targeted thoracic point-of-care ultrasound (TT-POCUS)-derived variables to predict did-not-finish (DNF) and first-treatment failure (FTF) outcomes in feedyard cattle diagnosed with bovine respiratory disease (BRD). First-treatment cattle were evaluated for both FTF and DNF outcomes, whereas chronic cattle were evaluated for DNF outcomes following enrollment. A prospective cohort observational study was conducted in cattle receiving either first treatment or chronic treatment for BRD. TT-POCUS examinations were performed at enrollment to assess pulmonary abnormalities, including lung consolidation, pleural irregularities, B-lines, and lesion scores. A total of 1422 first-treatment cattle and 150 chronic cattle were enrolled. Six supervised machine learning algorithms were trained using TT-POCUS variables and demographic data. For FTF prediction, logistic regression demonstrated the highest AUC (0.71). For first-treatment DNF, logistic regression again achieved the highest AUC (0.80). In chronic cattle, the perceptron model demonstrated the highest AUC (0.84). Overall, TT-POCUS-derived variables provided meaningful prognostic information and may support improved treatment decisions, antimicrobial stewardship, and animal welfare in feedyard cattle.

1. Introduction

Bovine respiratory disease (BRD) remains the most significant health challenge in feedyard cattle and is the leading cause of morbidity and mortality in the industry [1,2]. Bovine respiratory disease has been estimated to account for approximately 75% of morbidity and 50% to 70% of mortality during the finishing phase, highlighting its substantial effects on cattle health and feedyard productivity [3,4,5]. In addition to mortality losses, BRD negatively affects average daily gain, feed efficiency, days on feed, and carcass quality, resulting in considerable economic consequences for beef production systems [1,6]. Although reductions in BRD incidence can improve overall industry performance, economic benefits may not be evenly distributed across sectors, with feedyard operators often bearing a disproportionate share of disease-related costs [7].

The persistent burden of BRD reflects its multifactorial etiology, involving interactions among pathogens, host immunity, environmental stressors, and management practices [5,8]. Risk factors such as transportation stress, commingling, source variability, and nutritional stress contribute to disease development, yet no single factor fully explains disease occurrence [5]. Furthermore, subclinical disease and imperfect diagnostic methods complicate identification and management, emphasizing the need for improved objective diagnostic and prognostic tools.

Recent advances in data collection and computational methods have increased interest in predictive modeling and machine learning applications within livestock production systems. These approaches allow integration of multidimensional biological, environmental, and management data to identify complex relationships not readily captured using conventional statistical methods. Predictive modeling has been applied to diverse cattle production challenges, including thermal stress assessment, feed intake prediction, disease detection, reproductive management, and milk production forecasting [9,10,11,12,13,14].

Within beef production systems, predictive models have been evaluated to support management decisions such as metaphylaxis administration at arrival and mortality risk following BRD treatment [15,16]. These studies demonstrate that predictive tools may improve economic decision-making and intervention targeting. Predictive model performance depends heavily on data quality, predictor selection, and model structure. No single modeling methodology has consistently demonstrated universal superiority across all cattle health applications, suggesting model selection should align with the biological context, available data, and intended use [17,18].

Despite recent advances, comparatively few predictive modeling studies have focused on prognosis following BRD treatment initiation relative to arrival-risk classification or disease detection models. Existing prognostic approaches have primarily relied on demographic, management, or clinical variables, with limited incorporation of objective pulmonary imaging biomarkers such as targeted thoracic point-of-care ultrasound (TT-POCUS)-derived lesion characterization. Prognostic decision-making following BRD diagnosis remains clinically important because it may influence retreatment decisions, antimicrobial use, culling strategies, labor allocation, and animal welfare outcomes.

Development of accurate BRD prognostic models is challenging for several reasons. First, BRD diagnosis in field settings often relies on clinical signs that do not perfectly identify the true disease status of cattle. Consequently, some animals diagnosed with BRD may not actually have the disease (false positives), while others with BRD may go undetected (false negatives), reducing the reliability of data used for model development [18,19,20]. Second, class imbalance is common because severe outcomes such as mortality or non-recovery occur substantially less frequently than successful treatment responses. In imbalanced datasets, many classification algorithms tend to favor prediction of the majority class because doing so minimizes overall classification error. Consequently, fewer animals are classified as positive cases, increasing the likelihood of false-negative predictions and reducing model sensitivity despite apparently high overall accuracy [21]. Third, variation among production systems, treatment protocols, and outcome definitions may limit external validity and practical implementation. Collectively, these challenges indicate that improved biologically relevant predictors are needed to advance prognostic performance.

Thoracic ultrasonography has emerged as a valuable tool for evaluating pulmonary disease in cattle and offers strong potential for integration into predictive modeling frameworks. Unlike conventional clinical assessment methods, ultrasound provides objective, real-time visualization of lung consolidation, pleural abnormalities, interstitial changes, and lesion severity [22], and has outperformed traditional methods such as auscultation and rectal temperature for detection of pulmonary pathology in some settings [23]. Targeted thoracic point-of-care ultrasound (TT-POCUS) is a streamlined chute-side approach designed for rapid assessment under commercial feedyard conditions, and TT-POCUS-derived variables, including lung score, B-line patterns, lesion distribution, and consolidation severity, have been associated with clinically relevant outcomes such as first-treatment failure (FTF) and death or non-recovery (DNF) in cattle treated for BRD [24]. Because TT-POCUS can be performed rapidly during processing or hospital evaluation, it may provide a practical and objective tool for real-time risk stratification.

Given the substantial burden of BRD and limitations of current prognostic approaches, there is a need to evaluate whether TT-POCUS variables can improve prediction of treatment outcomes when combined with machine learning methods. Therefore, the objective of this study was to compare the performance of multiple supervised machine learning algorithms incorporating TT-POCUS-derived variables for prediction of FTF and DNF outcomes in feedyard cattle diagnosed with BRD. We hypothesized that models incorporating TT-POCUS variables would provide clinically useful prognostic discrimination and that model performance would vary across algorithm types and outcome definitions.

2. Materials and Methods

2.1. Experimental Design and Enrollment Criteria

This study was approved by the Institutional Animal Care and Use Committee (IACUC) at Kansas State University (Protocol #5060 and Protocol #5239).

This was a prospective cohort observational study conducted within the high-plains region of the United States. Cattle were enrolled from a single commercial feedyard between 2023 and 2025. Eligible animals included cattle presented to the feedyard hospital for their initial BRD treatment (first-treatment cohort) or cattle receiving a third or greater BRD treatment according to hospital protocol (chronic cohort). Cattle were identified as BRD positive by standard observational methods, and the reference is the field BRD diagnosis, which is susceptible to misclassification. Cattle evaluated at second treatment were not enrolled to preserve distinction between acute first-treatment and established chronic disease populations.

Animals enrolled at first treatment were not subsequently re-enrolled if they later progressed to chronic disease, ensuring each animal contributed only one observation to the dataset. At enrollment, demographic information, including sex, body weight, and days on feed, was recorded. Animals were excluded if a complete TT-POCUS examination could not be obtained because of time or animal handling constraints. Study outcomes included first-treatment failure (FTF) and death or non-recovery (DNF) within 60 days following enrollment, whereas chronic cattle were evaluated for DNF outcomes only during the 60-day follow-up period. Any animal that required BRD retreatment, was culled because of BRD, or died due to BRD within 60 days of enrollment was classified as FTF. For both first-treatment and chronic cohorts, animals that died or were culled because of BRD within the 60-day follow-up period were classified as DNF. Cattle were excluded from outcome analysis if they were unavailable for complete 60-day follow-up because of shipment for harvest or removal from the feedyard for reasons unrelated to BRD, including non-respiratory death. Cause of death and culling decisions were determined according to feedyard diagnostic records and attending staff assessments.

2.2. Targeted Thoracic Point-of-Care Ultrasound

The TT-POCUS method targeted the right caudodorsal lung lobe, beginning at the 11th intercostal space and progressing cranially to the 8th intercostal space, as previously described by Feitoza et al. (2025) [23,24,25,26]. Each intercostal space was scanned (iQ+Vet; 1 MHz to 10 MHz; Butterfly Network, Inc.; Burlington, MA, USA) dorsoventrally to ensure complete evaluation of the lung field. Examinations were performed chute-side and lasted approximately 60 s, and probe angulation and positioning were adjusted as needed to optimize visualization for pulmonary pathology. Real-time assessments of pulmonary abnormalities associated with BRD, including lung consolidation, pleural line irregularities, B-lines, A-lines, and ultrasound lung score (ULS), were recorded.

A-lines were defined as horizontal, repetitive hyperechoic reverberation artifacts arising from the pleural line and occurring at regular intervals deeper in the image, representing normal aerated lung and the absence of significant interstitial or alveolar fluid accumulation. B-lines were defined as vertical hyperechoic reverberation artifacts originating from the pleural line and extending to the far field without fading, consistent with increased pulmonary fluid or interstitial involvement. Coalescing or widened B-lines were interpreted as more severe interstitial pathology. Pleural line irregularities, described as a “moth-eaten” appearance termed “moth-sign”, represented disruption or unevenness of the normally smooth pleural surface and were associated with inflammatory or chronic pulmonary changes. Lung consolidation was defined as hypoechoic tissue-like regions within the lung parenchyma resulting from loss of normal aeration, often containing visible air bronchograms and consistent with bronchopneumonia.

Scoring was based on the most severe lesion identified across the scanned region. Ultrasound lung scores ranged from 1 to 5, where ULS 1 indicated minimal pathology characterized by fewer than 3 thin B-lines, consistent with normal or near-normal lung tissue. Although A-lines were recorded as a separate ultrasonographic variable, they were not incorporated into ultrasound lung score assignment. ULS 2 indicated the presence of 3 or more distinct B-lines, suggestive of mild interstitial involvement. ULS 3 indicated widening or coalescing B-lines consistent with more extensive interstitial pathology. ULS 4 represented pleural line irregularities or “moth-eaten” lesions indicating pleural disruption, while ULS 5 indicated lung consolidation consistent with severe pulmonary pathology.

2.3. Dataset Construction and Preprocessing

For model development, the dataset was divided into training (75%) and testing (25%) subsets using stratified random sampling based on outcome status (FTF or DNF, depending on the analysis cohort), with random_state = 123. The testing subset was withheld from model training and used for final model evaluation. As a supplemental internal validation analysis, repeated stratified five-fold cross-validation with five repeats and random_state = 123 was conducted to assess model performance stability across repeated resampling. Cross-validation results were interpreted separately from the final test-set performance estimates and were not used for hyperparameter tuning or model selection.

Candidate predictors were selected a priori based on biological relevance and included thoracic ultrasonography variables (lung score, A-line presence categorized as 0–2 and ≥3, B-line count categorized as 0–2 and ≥3, and moth-sign presence or absence) and animal-level characteristics (sex, body weight categorized as <272, 272–361, 362–453, and ≥453 kg, and days on feed categorized as 0–42, 43–71, and ≥72 days at time of assessment). A-line presence was included as a predictor in the chronic-treatment model but excluded from the first-treatment model because A-line observations were not consistently recorded during earlier study collections, resulting in substantial missingness. The missingness mechanism was not formally assessed; therefore, the data cannot be conclusively classified as missing completely at random (MCAR) or missing at random (MAR). Because data availability varied across collection periods, the missingness may be related to the timing of data collection. A-line presence was therefore excluded from the first-treatment models to avoid relying on a predictor with substantial missing data.

Data preprocessing was conducted within a unified machine learning pipeline to prevent information leakage and to prevent algorithms from accessing information outside the training set. Missing continuous variable values were imputed using median values, followed by standardization using z-score scaling. Missing categorical variable values were imputed using the most frequent category. Categorical predictors were encoded using one-hot encoding with the first category dropped to avoid collinearity. All preprocessing steps, including imputation, scaling, and encoding, were fit exclusively on the training dataset and subsequently applied to the testing dataset through the pipeline structure to ensure no information from the test set informed model training.

2.4. Model Development and Machine Learning Framework

Six supervised classification algorithms were evaluated: logistic regression (LR), decision tree (DT), random forest (RF), gradient boosting machine (GB), artificial neural network (NN), and perceptron (P). These models were selected to represent parametric, nonparametric, and ensemble approaches commonly used in biomedical and veterinary predictive modeling.

All models were implemented using scikit-learn. LR was specified with solver = “saga”, max_iter = 2000, and class_weight = “balanced”. RF was specified with 500 trees, a maximum depth of 20, class_weight = “balanced”, and random_state = 123. GB was specified with max_iter = 500, max_depth = 6, and random_state = 123. NN was specified as a multilayer perceptron with two hidden layers containing 50 neurons each, ReLU activation, max_iter = 500, and random_state = 123. P was specified with max_iter = 1000, tol = 0.001, and random_state = 123. DT was specified with max_depth = 10, class_weight = “balanced”, and random_state = 123. Inverse-frequency class weighting was therefore applied to LR, RF, and DT, whereas GB, NN, and P were not class weighted. The train–test split and applicable stochastic algorithms used a fixed random state of 123. All other model parameters were retained at their scikit-learn defaults. All analyses were conducted in Python 3.14.0 using standard scientific computing libraries, including scikit-learn (version 1.8.0), pandas, NumPy, and matplotlib.

The LR was included as a baseline parametric model because of its widespread use in clinical research and ease of interpretation. Logistic regression estimates the probability of a binary outcome and quantifies the direction and magnitude of predictor effects [27]. Although logistic regression is generally resistant to overfitting, it is limited in its ability to model nonlinear relationships unless these terms are explicitly specified.

The NN model was included to evaluate potentially nonlinear relationships and interactions among predictors. Complex interactions and nonlinear decision boundaries can be modeled through the hidden layers used in NN [27]. In contrast, the P model was included as a simple linear classification approach and provides a single-layer structure for binary classification [28].

Interpretability and flexibility with mixed data types motivated the selection of DT [29]. These models classify observations by recursively splitting predictor variables according to criteria such as information gain or Gini impurity, resulting in a rule-based structure. However, single DT models are often unstable and prone to overfitting.

To address these limitations, RF models were incorporated because they aggregate multiple decision trees built from bootstrap samples and random subsets of predictors. This ensemble strategy improves predictive accuracy, reduces variance, and captures nonlinear relationships and interactions [30]. Additional advantages of RF include out-of-bag error estimation and measures of variable importance.

Gradient boosting machines operate as sequential ensembles in which each new tree is trained to correct errors from previous trees. This approach enables flexible modeling of nonlinear effects, interactions, and threshold relationships [31]. Regularization strategies such as shrinkage and subsampling help mitigate overfitting. DT-based methods have further demonstrated strong performance in cattle health prediction, including outcomes related to bovine respiratory disease [21].

Class imbalance was partially addressed using inverse-frequency class weighting in applicable classifiers. Class weighting was applied to LR, RF, and DT models, whereas GB, NN, and P models were not class weighted. The P model was implemented using the scikit-learn Perceptron classifier with max_iter = 1000, tol = 0.001, and random_state = 123. All models were trained within a unified preprocessing pipeline designed to prevent information leakage. Final model performance was assessed on an independent testing dataset to evaluate generalizability and reduce overfitting.

2.5. Model Evaluation and Interpretation

Predictive performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and F1 score. These metrics were selected to assess overall discrimination and classification performance, particularly in the presence of class imbalance. The AUC measures a model’s ability to distinguish positive from negative outcomes across all thresholds by plotting sensitivity against 1 − specificity; values range from 0.5 (no discrimination) to 1.0 (perfect discrimination). Accuracy represents the proportion of correctly classified observations and is calculated as (TP (True Positive) + TN (True Negative)/(TP + TN + FP (False Positives) + FN (False Negatives)). Sensitivity, or recall, measures the proportion of true positives correctly identified and is calculated as TP/(TP + FN), while specificity measures the proportion of true negatives correctly identified and is calculated as TN/(TN + FP).

Positive predictive value, or precision, is the proportion of predicted positives that are truly positive and is calculated as TP/(TP + FP). Negative predictive value (NPV) is the proportion of predicted negatives that are truly negative and is calculated as TN/(TN + FN). The F1 score is the harmonic mean of PPV and sensitivity, calculated as 2 × (PPV × Sensitivity)/(PPV + Sensitivity), and is commonly used in imbalanced classification settings because it jointly summarizes precision and recall into a single metric.

Predicted class labels were generated using each model’s default classification rule through the predict () function in scikit-learn. For probabilistic classifiers, this generally corresponds to a default probability threshold of 0.50, whereas linear classifiers use their default decision boundary. Area under the receiver operating characteristic curve (AUC) was calculated separately using predicted probabilities or decision scores when available.

3. Results

3.1. First Treatment Failure

A total of 1422 first-treatment cattle were included in the FTF analysis, of which 476 (33.5%) experienced first-treatment failure within 60 days (Table 1). Descriptive characteristics of cattle classified as first-treatment success or FTF are summarized in Table 1.

Table 1. Descriptive statistics of 60-day first-treatment outcomes according to cattle demographics (sex, body weight, days on feed categories) and real-time chute-side evaluations (ultrasound lung score, B-line categories, and moth-sign). Outcomes included recovered, did not finish (DNF), first treatment success (FTS), and first treatment failure (FTF).

First Treatment 60-Day Outcome
	Recovered 1	DNF 2	FTS 3	FTF 4
Sex
Heifer	680	121	548	253
Steer	494	127	398	223
Days on feed at enrollment categories
0 to 42	814	171	653	332
43 to 71	156	32	115	73
>71	204	45	178	71
Body weight, average (SD), kg.	376.1	371.6	380.6	364.7
	(+/−84.8)	(+/−85.9)	(+/−86.4)	(+/−81.2)
Body weight, categories, kg.
<272	101	20	77	44
272 to 361	462	112	353	221
362 to 453	377	74	314	143
≥453	234	42	202	68
Real-time measurements 5
Ultrasound lung score, head
1	298	17	264	51
2	472	73	389	156
3	291	96	210	177
4	85	48	63	70
5	28	14	20	22
B-line category, head
0 to 2	700	170	571	299
≥3	474	78	375	177
Moth-sign, head 6
Absent	1083	248	876	408
Present	91	47	70	68

Animals that were not culled within 60 days post-evaluation, remained present with their cohort, and had not required re-treatment for BRD, ¹; animals that died or were culled because of BRD within 60 days after enrollment, ²; animals present with their cohort at 60 days post-evaluation and did not require re-treatment for BRD, ³; animals that required additional BRD treatment, were culled because of BRD, or died due to BRD within 60 days after initial treatment, ⁴; variables collected at chute side in real time, ⁵; moth-sign is noted when the pleural line shows irregularities like pleural thickening and sub-pleural consolidation (indentations), ⁶.

Model performance for prediction of FTF is summarized in Table 2. Among evaluated models, logistic regression demonstrated the highest overall discrimination (AUC = 0.71) with balanced sensitivity and specificity. The perceptron model showed comparable discrimination with the highest sensitivity, whereas neural network and gradient boosting approaches demonstrated higher specificity but substantially lower sensitivity. Decision tree models demonstrated the lowest overall discrimination.

Table 2. Descriptive statistics of first-treatment failure (FTF) model performance for Perceptron (P), Neural Network (NN), Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), and Gradient Boosting (GB). Performance metrics included area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and F1 score (F1).

	AUC	ACC%	Sn%	Sp%	PPV%	NPV%	F1%
P	0.68	63	77	56	46	83	58
NN	0.64	66	29	85	49	71	37
LR	0.71	66	67	65	48	80	57
RF	0.68	65	50	72	47	74	49
DT	0.63	59	56	60	42	73	48
GB	0.65	63	27	81	42	69	33

3.2. First-Treatment Did-Not-Finish

A total of 1422 first-treatment cattle were included in the DNF analysis, of which 248 (17.4%) were classified as DNF within 60 days (Table 1). Descriptive demographic and ultrasonographic characteristics for recovered and DNF cattle are summarized in Table 1.

Model performance for prediction of first-treatment DNF is presented in Table 3. Logistic regression demonstrated the highest overall discrimination (AUC = 0.80) and highest sensitivity, while neural network and gradient boosting models demonstrated the highest specificity and overall accuracy but markedly reduced sensitivity. Random forest models demonstrated moderate discrimination and more balanced classification performance across metrics. Negative predictive values remained consistently high across models, reflecting the relatively low prevalence of DNF within the study population.

Table 3. Descriptive statistics of first-treatment did not finish (DNF) model performance for Perceptron (P), Neural Network (NN), Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), and Gradient Boosting (GB). Performance metrics included area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and F1 score (F1).

	AUC	ACC%	Sn%	Sp%	PPV%	NPV%	F1%
P	0.59	71	36	79	26	85	30
NN	0.79	82	18	96	46	85	26
LR	0.80	70	81	68	35	94	48
RF	0.75	73	53	77	33	89	41
DT	0.68	67	63	68	29	90	40
GB	0.74	83	13	97	50	84	21

3.3. Chronic Treatment Did-Not-Finish

A total of 150 chronic-treatment cattle were included, consisting of 52 recovered and 98 DNF animals (65.3%) (Table 4). Descriptive demographic and ultrasonographic characteristics of recovered and DNF chronic-treatment cattle are presented in Table 4.

Table 4. Descriptive statistics of chronic treatment did not finish (DNF) outcome according to cattle demographics (sex, body weight, and days on feed categories) and real-time chute-side evaluations (ultrasound lung score, A-line count categories, B-line count categories, and moth-sign).

Chronic Treatment 60-Day Outcome
	Recovered 1	DNF 2
Sex
Heifer	31	60
Steer	21	38
Days on feed at enrollment categories
0 to 42	7	25
43 to 71	28	30
>71	17	43
Body weight, average (SD), kg.	389.8	366.3
	(+/−69.6)	(+/−87.2)
Body weight, categories, kg.
<272	3	15
272 to 361	14	32
362 to 453	26	33
≥453	9	18
Real-time measurements 3
Ultrasound lung score, head
1	5	1
2	11	16
3	25	33
4	9	37
5	2	11
A-line category, head
0 to 2	22	57
≥3	30	41
B-line category, head
0 to 2	13	3
≥3	39	95
Moth-sign, head 4
Absent	41	53
Present	11	45

Animals that were not culled within 60 days post-evaluation and present with their cohort and had not required re-treatment for BRD, ¹; animals that died or were culled because of BRD within 60 days after enrollment, ²; variables collected at chute side-real time, ³; moth-sign is noted when the pleural line shows irregularities like pleural thickening and sub-pleural consolidation (indentations), ⁴.

Model performance for chronic-treatment DNF prediction is presented in Table 5. The perceptron model had the highest observed discrimination in the chronic-treatment cohort (AUC = 0.84), whereas the random forest model achieved the highest sensitivity. Given the relatively small chronic-treatment cohort (n = 150), these differences should be interpreted cautiously and may be sensitive to the specific study sample. Logistic regression and neural network models demonstrated moderate-to-good discrimination with relatively balanced performance across classification metrics. In contrast, decision tree and gradient boosting approaches demonstrated lower overall discrimination.

Table 5. Descriptive statistics of chronic did not finish (DNF) model performance for Perceptron (P), Neural Network (NN), Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), and Gradient Boosting (GB). Performance metrics included area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and F1 score (F1).

	AUC	ACC%	Sn%	Sp%	PPV%	NPV%	F1%
P	0.840	76	80	69	83	64	82
NN	0.812	66	68	61	77	50	72
LR	0.779	68	64	77	84	52	72
RF	0.774	76	88	54	79	70	83
DT	0.675	58	52	69	77	42	62
GB	0.543	71	88	38	73	63	80

4. Discussion

This study evaluated multiple supervised machine learning algorithms incorporating TT-POCUS-derived variables to predict clinically relevant outcomes in feedyard cattle diagnosed with BRD. Predictive performance varied by outcome, with moderate discrimination observed for first-treatment failure (FTF) and moderate-to-good discrimination observed for did-not-finish (DNF) outcomes, particularly in chronic-treatment cattle. Across analyses, simpler approaches, including logistic regression (LR) and perceptron (P) models, frequently demonstrated performance comparable to or better than more complex nonlinear approaches. These findings suggest that TT-POCUS-derived pulmonary biomarkers may contribute to prognostic prediction while remaining compatible with operationally practical predictive models. However, the incremental prognostic value of TT-POCUS beyond demographic variables could not be determined because a demographic-only model was not evaluated.

Several limitations should be considered when interpreting these findings. Cattle were enrolled from a single commercial High Plains feedyard under field conditions, and the models may reflect population-specific characteristics, including cattle sourcing, breed composition, pathogen exposure, management practices, and treatment protocols. Potential outcome misclassification associated with field-based BRD diagnosis is another limitation. Although a held-out test set was used for model evaluation, this represents internal evaluation within the study population rather than external validation in an independent operation. Therefore, the generalizability of these models to other feedyards and production settings remains uncertain, and evaluation in independent cattle populations is needed before broader application.

Predictive performance for FTF was lower than for DNF outcomes, suggesting that early retreatment risk may be influenced by factors extending beyond pulmonary lesion severity alone. Differences were also observed between first-treatment and chronic-treatment DNF outcomes. First-treatment DNF models demonstrated good discrimination, with logistic regression achieving the highest AUC (0.80). However, performance was characterized by high negative predictive values (84–94%) and lower positive predictive values (26–50%) across algorithms, a pattern consistent with the relatively low prevalence of DNF in this population. In the chronic-treatment cohort, higher observed discrimination was achieved, with the perceptron model demonstrating the highest AUC (0.84), while several models demonstrated greater sensitivity (80–88%) and positive predictive values (73–84%). These findings may reflect differences in the characteristics and severity of disease between first-treatment and chronically treated cattle; however, this explanation requires further evaluation. The performance of simpler models, including logistic regression and the perceptron, suggests that complex nonlinear approaches did not consistently provide better discrimination using the predictors evaluated in this study.

The chronic-treatment analysis was based on a relatively small cohort of 150 cattle, and evaluation of multiple algorithms in this dataset raises the possibility of overfitting and unstable performance estimates. Accordingly, the AUC of 0.84 observed for the perceptron should be considered exploratory rather than definitive evidence of superior model performance. Validation in a larger, independent chronic-treatment population is needed to determine whether these findings can be reproduced.

This observation aligns with previous BRD predictive modeling studies, where increased algorithmic complexity has not consistently translated into improved predictive performance when using structured clinical datasets [21,32,33]. Although ensemble and neural network approaches are capable of modeling nonlinear interactions, they may also be more sensitive to noise, feature redundancy, and class imbalance commonly present in field-derived veterinary datasets [18]. Similarly, Amrine et al. (2014) reported substantial variation in classifier performance across feedyards and concluded that no single model consistently outperformed all others [34]. This likely reflects biological and management heterogeneity among operations, including differences in cattle sourcing, pathogen exposure, antimicrobial protocols, environmental stressors, and BRD case definitions. Consequently, predictor-outcome relationships identified in one population may not generalize consistently across operations, emphasizing the importance of local validation and model recalibration.

The lower discrimination observed for FTF prediction, relative to other outcomes, likely reflects the multifactorial nature of early treatment response. FTF is influenced not only by pulmonary pathology but also by antimicrobial selection, timing of intervention, host immune response, pathogen susceptibility, and environmental stressors [18]. As such, TT-POCUS provides important but incomplete information regarding treatment response dynamics. The incremental prognostic value of TT-POCUS beyond demographic variables cannot be quantified because a demographic-only model was not evaluated in this study. Therefore, the observed model performance should not be interpreted as evidence that incorporating TT-POCUS improves prediction beyond demographic information alone. Future studies should compare models incorporating demographic variables alone with models incorporating both demographic and TT-POCUS-derived variables to determine the additional predictive value of thoracic ultrasonography. This pattern is further supported by the stronger discrimination observed for DNF compared with FTF among first-treatment cattle, suggesting that severe outcomes may be more directly associated with measurable pulmonary pathology than short-term retreatment risk.

The relative importance of sensitivity versus specificity depends on the intended clinical application of the model. For prognostic BRD decision-making, sensitivity may be particularly important because false-negative predictions result in cattle at high risk of FTF or DNF being incorrectly classified as low risk. Failure to identify these animals may delay additional monitoring or intervention and could contribute to increased mortality, chronicity, welfare concerns, and economic losses. Conversely, false-positive classifications may increase antimicrobial use, labor requirements, and hospital resource utilization. Therefore, optimal model selection and classification thresholds should be determined according to the relative consequences of false-negative and false-positive errors within a given management system. From a management perspective, improved sensitivity in chronic cattle is particularly relevant because these animals often require repeated antimicrobial therapy, prolonged hospitalization, altered marketing strategies, or culling decisions.

Research previously demonstrated that TT-POCUS abnormalities, including elevated lung scores, increased B-line patterns, and moth-sign lesions, were associated with adverse 60-day outcomes in feedyard cattle [24]. The present study extends those findings by incorporating these lesion-based variables into supervised machine learning frameworks for outcome prediction.

When considered alongside prior predictive modeling efforts in BRD, these findings support further evaluation of TT-POCUS-derived variables as candidate predictors of prognostic outcomes. However, direct comparisons with models based on conventional data sources alone were not performed in the present study. Heinen et al. (2023) evaluated 96,382 BRD-treated cattle using routine feedyard variables and reported a median AUC of 0.675, with boosted decision trees performing best (AUC = 0.738) [21]. In contrast, the present study achieved an AUC of 0.80 using logistic regression, suggesting that ultrasound-derived pulmonary biomarkers may provide clinically meaningful prognostic information beyond demographic and management variables alone.

Future integration of TT-POCUS-derived variables with existing demographic-based prognostic models, such as those described by Heinen et al., could be explored using either parallel or serial testing frameworks. In a parallel framework, sensitivity may improve by identifying cattle flagged by either model, whereas serial application may improve specificity by restricting high-risk classification to cattle identified by both approaches. Future work should evaluate how these combined strategies affect predictive values and clinical utility under commercial conditions.

These findings are consistent with prior chute-side diagnostic technologies such as Whisper^® lung scoring, where incorporation of standardized respiratory auscultation substantially improved prognostic discrimination for BRD mortality compared with rectal temperature alone [35]. In that study, inclusion of Whisper^® lung scores increased ROC AUC from 0.50 to 0.64, while combining Whisper^® scoring with rectal temperature further improved AUC to 0.69. Whisper^® scoring also markedly reduced false-negative classifications relative to fever alone, indicating improved sensitivity for identifying cattle at risk of death. Similar to Whisper^®, TT-POCUS provides structured lesion-based pulmonary assessment rather than reliance solely on rectal temperature or subjective clinical appraisal. The comparatively stronger performance observed in the present study, particularly for DNF outcomes where LR achieved an AUC of 0.80, suggests that direct ultrasonographic characterization of pulmonary pathology may provide additional prognostic information beyond traditional chute-side clinical assessment methods.

Amrine et al. (2014) reported that predictive accuracy varied across 23 feedyards and that locally derived models often outperformed combined datasets, highlighting the importance of biological and management context [34]. The variability in model performance reported may partially explain why simpler and more interpretable models generalized effectively within the present feedyard population. Furthermore, objective lesion-based predictors such as TT-POCUS may reduce some variability associated with subjective clinical assessment because TT-POCUS provides standardized lesion-based pulmonary evaluation rather than relying exclusively on visual appraisal of clinical signs.

The consistently high negative predictive values across models are particularly relevant for clinical application. In low-prevalence outcomes such as DNF, high NPV indicates strong reliability in identifying cattle unlikely to experience poor outcomes, which may support reduced intervention intensity in animals at low risk of FTF or DNF and more focused allocation of hospital resources toward higher-risk cattle [18].

The improved predictive performance observed in chronic cattle supports the concept that TT-POCUS may provide greater prognostic discrimination when pulmonary pathology is more established. In these animals, TT-POCUS abnormalities may be more pronounced and therefore more strongly associated with clinical outcomes. Prior TT-POCUS research has similarly demonstrated that larger areas of lung consolidation, higher ultrasound lung scores, pleural irregularities, and extensive comet-tail artifacts are associated with poorer clinical outcomes and increased mortality risk in feedyard cattle [24,25].

Earlier identification of cattle unlikely to recover may also improve financial outcomes by allowing earlier culling decisions before prolonged feed, treatment, and labor costs accumulate. Previous economic analyses have shown that delayed culling of chronically ill cattle can substantially increase net return losses compared to earlier removal decisions [36].

The findings of this study suggest that integrating TT-POCUS-derived variables with machine learning approaches may provide practical value for prognostic decision-making in feedyard cattle identified for BRD treatment. The practical value of TT-POCUS-based prediction lies in its ability to convert objective lesion-based information into actionable management decisions at the time of evaluation.

In practical feedyard settings, these tools could assist veterinarians and personnel in prioritizing monitoring intensity, retreatment decisions, hospital pen allocation, and antimicrobial use. Because LR and P models demonstrated competitive predictive performance despite their relative simplicity, clinically useful implementation may not require highly complex computational frameworks. This may facilitate integration into practical feedyard decision-support systems with reduced computational burden and greater interpretability for veterinary personnel.

The results also reinforce that predictive model utility should be matched to the intended management objective. Models with higher sensitivity may be preferred when the goal is early identification of chronic cattle, whereas models with higher specificity may be advantageous when minimizing unnecessary intervention is the priority. This principle has been emphasized in prior BRD predictive modeling reviews, where cutoff selection should depend on the relative cost of false positives and false negatives rather than a single universal threshold [18].

Several limitations should be considered when interpreting these findings. First, the study population was derived from a single commercial feedyard, which may limit generalizability to other regions, management systems, cattle populations, and BRD case definitions. Previous work has demonstrated yard-specific models often outperform pooled models, highlighting substantial between-operation heterogeneity [21,34].

Second, BRD diagnosis and outcome classification under field conditions are inherently imperfect and may introduce misclassification that affects both model training and evaluation. In the absence of a definitive gold standard for BRD, this limitation is common across predictive modeling studies [18,32].

Incomplete availability of some predictors, including full A-line data for first-treatment cattle, limited uniform inclusion across analyses. Additional clinical, treatment, environmental, behavioral, and pathogen-related variables may improve future model performance, as suggested by prior work emphasizing precision livestock data streams and multimodal integration [18,32].

Conclusions drawn from the evaluation of the chronic cattle should be carefully interpreted due to the small sample size. Only 150 heads were available in this dataset, and this is very small for machine learning techniques, which could result in model instability, especially in complex algorithms.

Although internal train–test validation was performed, external validation in independent feedyard populations is still required prior to clinical implementation. In addition, default classification thresholds were used for binary outcomes, and future work incorporating threshold optimization based on economic or antimicrobial stewardship objectives may improve practical application.

Future research should evaluate integration of TT-POCUS-derived variables with broader clinical, management, and environmental datasets to improve predictive performance. External validation across multiple feedyards and optimization of classification thresholds based on economic and stewardship priorities will be necessary before widespread implementation.

5. Conclusions

Overall, this study demonstrates that TT-POCUS-derived variables provide clinically meaningful prognostic information for BRD outcomes in feedyard cattle when incorporated into supervised machine learning models. Predictive performance was moderate for first-treatment failure (FTF) and higher for death or non-recovery (DNF), suggesting that ultrasound-based measures may be more effective for identifying cattle at risk of poor longer-term outcomes than for predicting early treatment response. Among the evaluated approaches, simpler methods such as logistic regression outperformed more complex nonlinear and ensemble-based algorithms, indicating that increased model complexity did not consistently translate into improved predictive utility within this dataset.

Collectively, these findings support TT-POCUS as a practical, objective source of chute-side information that can enhance risk stratification when combined with routinely collected animal-level data. Data included in the current study were from a single feedyard, and further evaluation in varied production environments would be important to validate external validity. In a production context, this type of prognostic information may assist veterinarians and feedyard personnel in refining treatment decisions, targeting monitoring intensity, and improving allocation of resources for high-risk cattle. These models require further refinement and external validation across diverse feedyard populations before widespread implementation.