Machine Learning Model for an App-Based Tool to Assist With the Diagnosis of Canine Atopic Dermatitis

Abstract

Background: Canine atopic dermatitis (cAD) is a chronic condition requiring life-long management. Accurate diagnosis can be challenging, with no reliable diagnostic test. Objectives: This study aimed to generate a simple diagnostic model for cAD. Materials and Methods: Machine learning by Random Forest was applied to metadata from a prospective dataset of clinical cases definitively diagnosed with cAD alone or another dermatosis. The dataset underwent a division of 67% for training and 33% for testing, with the model being trained via stratified K-fold cross-validation on the former portion, while performance assessment was conducted on the latter portion. Nine referral clinicians across four European countries contributed 645 cases. Results: Modelling confirmed the value of the four tested metadata on a dog's history and reduced the initial 15 lesion locations tested to three. Metadata for the final model were: predisposed breed (any from a list of 31), predominantly indoor life, dermatitis onset age between 6 months and 3 years, dermatitis chronic, recurrent or a permanent background. Lesion locations were axilla, inguinal and other. Diagnostic prediction was 95% sensitive and 84% specific. Conclusions and Clinical Relevance: This model is a relevant prototype for an app-based tool to support general practitioners in the diagnosis of cAD alongside existing tests. It has high sensitivity and specificity based on four questions and three lesion locations obtained from a standard history and clinical examination.