(EN11) Machine Learning Algorithms as a Screening Tool for Hypercortisolism in Dogs

Abstract: Background Hypercortisolism (HC) is a common endocrine disorder in dogs that can significantly reduce the quality of life in affected animals. However, diagnosing HC is often challenging due to its variable clinical presentation, making it difficult to identify suitable candidates for further diagnostic tests for HC could be difficult. Hypothesis/Objectives To develop machine learning algorithms to assist in the diagnosis of HC using readily available routine tests. Animals 153 client-owned control dogs (suspected of HC but later excluded based on hormonal test results) and 152 dogs diagnosed with confirmed HC. Methods This case-control study utilized routine laboratory data, including complete blood count (CBC), serum chemistry panel, and urinalysis parameters such as urine specific gravity (USG) and urine protein-to-creatinine ratio (UPC). A boosted tree algorithm (Gradient Boosting) was trained using 80% of the collected data, while the remaining 20% was used for test data performance evaluation. Results The developed model demonstrated an accuracy of 88.5% (95% CI, 80.5%–96.5%), a sensitivity of 83.3% (95% CI, 70.7%–96.7%), a specificity of 93.5% (95% CI, 84.9%–100%), and an area under the receiver operating characteristic curve (AUC) of 0.912 (95% CI, 0.835–0.988), indicating excellent discriminatory ability between HC and controls. A user-friendly graphical interface was developed to enable practitioners to apply this tool for HC screening easily. Conclusions and Clinical Importance The developed machine learning algorithms have the potential to improve diagnostic efficiency and owner satisfaction by identifying appropriate candidates for HC testing and reducing unnecessary diagnostic procedures.