A machine learning-based web application aids Kawasaki disease diagnosis and predicts its outcomes

A machine learning-based web application aids Kawasaki disease diagnosis and predicts its outcomes. Download PDF Abstract Unless Kawasaki disease (KD) is promptly treated with intravenous immunoglobulin (IVIG), patients may develop IVIG-unresponsiveness and aneurysm. We analysed 628 KD and 155 KD-like cases to develop machine learning models for having KD (i.e., being diagnosed with KD), IVIG-unresponsiveness and aneurysm formation.

We applied LASSO logistic regression to illness-day-indexed data incorporating family/previous histories and laboratory variables (e.g., neutrophils, total bilirubin, C-reactive protein, and AST). Optional inputs included BNP, IgG, fibrinogen, and echocardiographic Z scores. On Day 4, all models achieved AUCs > 0.7 in the test dataset: having KD (0.779), IVIG-unresponsiveness (0.772), and aneurysm formation (0.710).

On Day 5, the model achieved a positive predictive value > 90% for having KD. When combined with predictions of unresponsiveness or aneurysm, 90% of complicated KD cases were identified. These models were implemented in a web-based and intranet-integrated application, enabling early diagnosis and risk stratification of KD across resource-limited and resource-rich settings.

Medical Disclaimer: This article is for informational purposes only and does not constitute medical advice. Always consult with qualified healthcare professionals for medical decisions and treatment options.