AI Tool Identifies Which Young Heart Patients Will Fail Standard Treatment Before Therapy Begins
Machine learning model using eight routine clinical variables predicts resistance to immunoglobulin therapy in Kawasaki disease with nearly 80% accuracy, outperforming all existing risk scores.
Researchers have developed a machine learning system that can identify which children with Kawasaki disease—the leading cause of acquired heart disease in pediatrics—will not respond to standard treatment, potentially enabling clinicians to intensify therapy before complications develop.
The support vector machine model, created and validated by investigators at Children's Hospital of Soochow University, achieved an area under the curve of 0.782 using eight routine clinical variables available at hospital admission. The tool significantly outperformed five established scoring systems currently used to assess treatment resistance risk, according to findings published in the World Journal of Pediatrics.
Addressing a Critical Gap in Pediatric Cardiology
Kawasaki disease causes coronary artery lesions in 20-25% of untreated patients, but standard intravenous immunoglobulin (IVIG) therapy prevents most cardiac complications. The challenge lies in the 10-20% of children who prove resistant to IVIG—a population at substantially higher risk for heart damage that requires more aggressive upfront treatment.
The model was developed using data from 2,371 children treated between December 2018 and June 2024, of whom 12.9% were IVIG-resistant. The eight predictors include coronary artery lesions, C-reactive protein, age at presentation, neutrophil percentage, white blood cell count, total cholesterol, platelet count, and serum albumin.
Validation across 894 additional patients from three cohorts—including 253 prospectively enrolled children—demonstrated consistent performance, with AUC values ranging from 0.746 to 0.799. The model achieved sensitivity of 0.735 and specificity of 0.721, with calibration scores indicating strong agreement between predicted and observed resistance rates.
Explainability and Early Risk Stratification
The researchers tested twelve different machine learning algorithms, with gradient boosting, AdaBoost, random forest, and support vector machine emerging as top performers. They employed SHAP (SHapley Additive exPlanations) values to make the model interpretable, addressing the "black box" problem that often limits clinical adoption of AI tools. Albumin showed the highest interaction strength among variables.
The team also developed an alternative five-variable model that excludes coronary artery lesions, achieving an AUC of 0.762—not statistically different from the full model. This version enables risk stratification using only blood tests and age, before echocardiography results are available, potentially accelerating clinical decision-making in emergency settings.
The investigators deployed the tool as an interactive web-based calculator built on the Shiny platform, providing clinicians with immediate access to predictions at the point of care.
Validation Standards and Limitations
The study followed TRIPOD+AI reporting guidelines for clinical prediction models. However, a PROBAST+AI assessment identified high risk of bias in participants, predictors, and analysis domains due to the retrospective design of the development and two external validation cohorts.
Age-related heterogeneity emerged in the analysis: infants 0-5 months showed only neutrophil percentage and total cholesterol differing between resistant and responsive patients, while children 3-4 years showed seven of eight features as significant. This suggests the model's performance may vary across age groups.
The optimal probability threshold of 14.9% yielded sensitivity of 0.648 and specificity of 0.606, indicating that while the tool improves upon existing scores, it will still miss some resistant cases and flag some responsive patients as high-risk.
Clinical Implementation Pathway
For healthcare providers, the validated tool offers a data-driven approach to identifying high-risk patients who may benefit from intensified first-line therapy—such as IVIG plus corticosteroids or immunosuppressants—rather than standard IVIG monotherapy. Early identification could prevent coronary complications in resistant cases while avoiding unnecessary treatment intensification in the 80-90% of children who respond to standard therapy.
The multi-site validation, including prospective real-world testing, represents a standard that regulators and payers increasingly expect from clinical AI applications, potentially smoothing the pathway from research to routine implementation.
Sources
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.