Written by Kimiko Dunbar
Spoon Feed
In this retrospective multi-center study, a machine learning model trained to predict Kawasaki Disease performed with over 89% accuracy.
We all need, machine-learning, to lean on…
Kawasaki disease (KD) is a relatively common but often diagnostically challenging childhood disease as there is no specific test for diagnosis. Further, later diagnosis is associated with higher rates of complication – i.e. coronary artery aneurysm.
This retrospective multicenter study externally validated “Kawasaki Match,” a machine learning tool designed to distinguish KD from other febrile illnesses. Data input into the model included the 5 clinical KD criteria, age, and the following laboratory data: WBC, hemoglobin (age adjusted), platelets, percent neutrophils, percent monocytes, percent eosinophils, ANC, ESR, CRP, ALT, albumin, and sodium.
Trained on these clinical and lab data, the model achieved: AUC 0.97, accuracy 0.92. sensitivity 0.94, specificity 0.88, PPV 0.9. NPV 0.93 on internal validation. External validation across 3 children’s hospitals achieved >91.4% accuracy and AUC 0.98. Sensitivity remained ≥89% across KD subgroups. Limitations include diagnostic imprecision without a gold standard and overrepresentation of KD cases. Given the nature of the study, the prevalence of KD was far higher than that which presents to the ED. The model shows promise as a decision support tool to aid timely KD diagnosis; however, we need some prospective studies in the clinical environment to better assess its on the ground utility.
How does this change my practice?
This machine learning tool seems promising. Given the overlap of KD symptoms with other illnesses, such as adenovirus, I would love to be able to plug patient information into a risk calculator to better inform my level of concern. However, we still need more information on the application of the model in a prospective way. A trial is currently underway at Rady Children’s Hospital in San Diego, and I look forward to seeing those results.
Source
External Validation of a Machine Learning Model to Diagnose Kawasaki Disease. J Pediatr. 2025 Mar 21;282:114543. doi: 10.1016/j.jpeds.2025.114543. Epub ahead of print. PMID: 40122277
