Machine learning classification of dilated cardiomyopathy severity in commercial broiler chickens

Dilated cardiomyopathy (DCM) is a significant cardiovascular disorder in commercial broiler chickens associated with congestive heart failure, ascites, and sudden death, leading to substantial economic losses in poultry production. Objective and accurate identification of DCM severity remains challenging due to reliance on subjective postmortem scoring systems. This study evaluated seven machine learning algorithms for classifying DCM severity into three ordinal categories (Score 1 = normal, Score 2 = moderate, Score 3 = severe) using 15 features comprising 6 cardiac morphometric measurements and 9 physiological and production parameters from 430 male Cobb 500 broiler chickens. Models were trained using 5-fold cross-validation with hyperparameter tuning via grid search, and stability was assessed through repeated 10 × 5-fold stratified cross-validation. XGBoost, Ridge regression, and Elastic Net achieved the highest test accuracy of 74.7% (95% CI: 64.2 to 83.4%), with macro AUC values of 0.85 for all three models. No statistically significant differences were observed between model performances (McNemar's test, all P > 0.05). Morphometric features, particularly heart width (w1) and the width-to-length ratio (w1l1), contributed approximately 62% of total feature importance, while physiological parameters contributed 38%. SHapley Additive exPlanations (SHAP) analysis confirmed that w1 was the most influential predictor across all severity classes. Score 3 (severe DCM) classification was limited by small sample size (n = 29, 6.7%). On the held-out test set, five of seven models achieved Score 3 sensitivity of 33.3% (2 of 6 cases); however, bootstrap analysis on cross-validated predictions yielded a lower mean sensitivity of 6.9% (95% CI: 0.0 to 17.2%), reflecting the high instability of this estimate. A supplementary binary classification (Normal vs. DCM) using repeated cross-validation yielded mean accuracies up to 79.0% with AUC up to 0.846. These findings demonstrate that machine learning approaches can provide objective and repeatable support for DCM severity assessment, with cardiac morphometric measurements providing the strongest discriminative signal. However, predictive performance was weakest and least stable for the severe class (Score 3), owing to its limited representation (6.7% of birds). Future work should focus on increasing severe DCM sample representation and integrating imaging-based features for improved classification.