UC San Diego

Kawasaki disease (KD), the most common cause of acquired heart disease in children, can be easily missed as it shares clinical findings with other pediatric illnesses, leading to risk of myocardial infarction or death. KD remains a clinical diagnosis for which there is no diagnostic test, yet there are classic findings on exam that can be captured in a photograph. This study aimed to develop a deep convolutional neural network, KD-CNN, to differentiate photographs of KD clinical signs from those of other pediatric illnesses. To create the dataset, we used an innovative combination of crowdsourcing images and downloading from public domains on the Internet. KD-CNN was then pretrained using transfer learning from VGG-16 and fine-tuned on the KD dataset, and methods to compensate for limited data were explored to improve model performance and generalizability. KD-CNN achieved a median AUC of 0.90 (IQR 0.10 from tenfold cross validation), with a sensitivity of 0.80 (IQR 0.18) and specificity of 0.85 (IQR 0.19) to distinguish between children with and without clinical manifestations of KD. KD-CNN is a novel application of CNN in medicine, with the potential to assist clinicians in differentiating KD from other pediatric illnesses and thus reduce KD morbidity and mortality.