Could a day-four model flag Kawasaki disease and complications earlier?
A Japanese hospital study separated 628 Kawasaki cases from 155 similar illnesses and predicted treatment resistance and aneurysm risk—but external and prospective validation are still missing.
The short version
- What happened
- Researchers analysed 628 children diagnosed with Kawasaki disease and 155 children with similar presentations, using LASSO logistic regression on history and laboratory data.
- Who it affects
- Families could benefit from faster escalation when Kawasaki disease is uncertain, but false reassurance or over-treatment would be serious.
- What remains unproven
- Single-hospital retrospective dataset
Keep in mind
Single-hospital retrospective dataset
No independent external or prospective validation
Discrimination does not prove calibration or better outcomes
Source analysis · 1 linked source. Primary analysed source date: 24 October 2026.
Read the full report · 2 min The evidence, methods, implications and limitations
Three models addressed diagnosis, treatment response and aneurysm risk
A team at Fukuoka Tokushukai Hospital organised records by illness day and trained LASSO logistic-regression models using family and previous history plus laboratory measures such as neutrophils, bilirubin, C-reactive protein and AST. Optional inputs included BNP, IgG, fibrinogen and echocardiographic Z scores. The dataset contained 628 Kawasaki cases and 155 Kawasaki-like cases.
On day four, held-out test AUCs were 0.779 for Kawasaki diagnosis, 0.772 for resistance to intravenous immunoglobulin and 0.710 for aneurysm formation. On day five, the diagnosis model's positive predictive value exceeded 90%. Combining complication predictions identified 90% of complicated cases in the reported data.[1]
An application does not make the model clinically validated
Earlier support could matter because delayed treatment raises coronary-complication risk, and common clinical fields may be practical when optional tests are unavailable. The team implemented web and intranet versions, but deployment is not a trial of benefit.
The cohort came from one Japanese hospital. AUC also hides threshold trade-offs: services need to know how many children are missed and how many receive unnecessary escalation. External validation, calibration, subgroup analysis and a prospective clinician-impact study are the next tests.[1]
The distinction that matters
What the tool predicts
The application presents three probabilities rather than one definitive diagnosis.
Diagnosis
Compare Kawasaki with similar presentations
Day-four AUC 0.779
Treatment
Estimate IVIG non-response
Day-four AUC 0.772
Aneurysm
Estimate coronary-complication risk
Day-four AUC 0.710