Using machine learning to diagnose Addison's disease in dogs

Abstract

Canine hypoadrenocorticism (CHA) is a life-threatening condition that affects approximately 3 of 1,000 dogs. It has a wide array of clinical signs and is known to mimic other disease processes, including kidney and gastrointestinal diseases, creating a diagnostic challenge. Because CHA can be fatal if not appropriately treated, there is risk to the patient if the condition is not diagnosed. However, the prognosis is excellent with appropriate therapy. A major hurdle to diagnosing CHA is the lack of awareness and low index of suspicion. Once suspected, the application and interpretation of conclusive diagnostic tests is relatively straight forward. In this study, machine learning methods were employed to aid in the diagnosis of CHA using routinely collected screening diagnostics (complete blood count and serum chemistry panel). These data were collected for 908 control dogs (suspected to have CHA, but disease ruled out) and 133 dogs with confirmed CHA. A boosted tree algorithm (AdaBoost) was trained with 80% of the collected data, and 20% was then utilized as test data to assess performance. Algorithm learning was demonstrated as the training set was increased from 0 to 600 dogs. The developed algorithm model has a sensitivity of 96.3% (95% CI, 81.7%-99.8%), specificity of 97.2% (95% CI, 93.7%-98.8%), and an area under the receiver operator characteristic curve of 0.994 (95% CI, 0.984-0.999), and it outperforms other screening methods including logistic regression analysis. An easy-to-use graphical interface allows the practitioner to easily implement this technology to screen for CHA leading to improved outcomes for patients and owners.