A deep convolutional neural network for Kawasaki disease diagnosis

Area of Science:

• Pediatric Cardiology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Kawasaki disease (KD) is a leading cause of acquired heart disease in children.
• KD diagnosis is challenging due to overlapping symptoms with other pediatric conditions, risking delayed treatment and severe outcomes like myocardial infarction.
• Currently, KD diagnosis relies on clinical evaluation as no definitive diagnostic test exists.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (KD-CNN) for differentiating clinical signs of KD from other pediatric illnesses using photographs.
• To assess the performance of KD-CNN in distinguishing between children with and without clinical manifestations of KD.

Main Methods:

• Utilized a hybrid approach for dataset creation, combining crowdsourced images and publicly available internet data.
• Employed transfer learning by pretraining a VGG-16 model and subsequently fine-tuning it on the KD dataset.
• Explored data augmentation techniques to enhance model performance and generalizability with limited data.

Main Results:

• KD-CNN achieved a median Area Under the Curve (AUC) of 0.90 (IQR 0.10) across tenfold cross-validation.
• The model demonstrated a median sensitivity of 0.80 (IQR 0.18) and a median specificity of 0.85 (IQR 0.19).
• These results indicate a strong ability to distinguish between children with and without clinical signs of KD.

Conclusions:

• KD-CNN represents a novel application of convolutional neural networks in medical diagnostics.
• The AI tool shows significant potential to assist clinicians in the early and accurate differentiation of Kawasaki disease.
• Successful implementation of KD-CNN could lead to reduced morbidity and mortality associated with KD.