Predicting the likelihood of hypoadrenocorticism in dogs using signalment and routine laboratory results with an ensemble machine learning predictive model

Background The adrenocorticotropic hormone stimulation test (ACTHst), required for diagnosing hypoadrenocorticism in dogs (canine hypoadrenocorticism, CHA), is limited by the cost and availability of synthetic adrenocorticotropic hormone, and results are subject to delays. Objectives Develop and validate a machine learning model that predicts the probability of CHA using signalment and routine laboratory test results. Animals Sixty-eight confirmed and untreated CHA dogs, and 504 control dogs (CHA suspected but ultimately excluded). Methods Cross-sectional multicenter study. Dogs in which CHA was confirmed or excluded by resting cortisol measurement or an ACTHst were identified from medical records of 5 veterinary referral hospitals. Data from 4/5 institutions were used to train a parallel random forest algorithm (parRF), the output of which was the predicted probability of CHA. Model performance was assessed on training set data (internal validation) and a different population (external validation from the fifth hospital). Results The parRF accurately predicted CHA in internal validation (area under the receiver-operating characteristic curve [ROC AUC]: 0.998; 95% CI, 0.996-0.999, 50%-predicted-probability sensitivity: 97.3%; 95% CI, 93.6-98.6, and specificity: 98.6%; 95% CI, 95.9-99.5). Three decision thresholds were determined: predicted probabilities of 10%, 50%, and 80% to optimize sensitivity, accuracy, and specificity, respectively. The model accuracy was confirmed using external data: ROC AUC: 0.942 (95% CI, 0.853-1.0); sensitivity: 94.1% (95% CI, 71.3-99.9), 58.8% (95% CI, 32.9-81.6), 41.2% (95% CI, 18.4-67.1); specificity: 93.9% (95% CI, 85.2-98.3), 100% (95% CI, 94.6-100), 100% (95% CI, 94.6-100), at 10%, 50%, and 80% decision thresholds, respectively. Conclusions and clinical importance Although the parRF should be further externally validated, it shows promise to help veterinarians diagnose CHA.