Kawasaki disease is an acute febrile vasculitis that predominantly affects young children and, if untreated, leads to coronary artery lesions in approximately 20 to 25 percent of patients, with the potential for luminal narrowing, myocardial ischemia, and tissue necrosis. High-dose intravenous immunoglobulin (IVIG) is the standard initial treatment, yet 10 to 20 percent of children are resistant to it and remain at heightened risk of severe cardiac complications. Early identification of these resistant cases is critical because intensified regimens, such as IVIG combined with corticosteroids or immunosuppressants, are recommended for high-risk patients. To meet this need, the research team built a support vector machine (SVM) model trained on routine clinical and laboratory data and tested it across multiple independent patient cohorts.

The study, published in World Journal of Pediatrics, drew on electronic medical records from Children’s Hospital of Soochow University in Suzhou, where a retrospective development cohort of 2,371 children with Kawasaki disease treated between December 2018 and June 2024 was assembled. External validation was performed on 443 patients from Fuzhou and 198 from Yangzhou, and a prospective cohort of 253 patients treated in Suzhou between July and December 2024 provided an additional real-world test. Among the development cohort, 12.9 percent were IVIG-resistant, consistent with previous reports from Chinese populations. The median age was 26.0 months, and 60.9 percent of patients were male.

The team evaluated twelve machine learning algorithms, including gradient boosting, adaptive boosting, random forest, neural networks, and regularized generalized linear models, using repeated 10-fold cross-validation for training and hyperparameter tuning. Gradient boosting achieved an internal AUC of 0.773, followed closely by AdaBoost at 0.765 and random forest at 0.756, with the SVM at 0.750. After considering both discrimination and balanced sensitivity and specificity, the researchers narrowed the field to three algorithms and then applied SHapley Additive exPlanations (SHAP) values, an explainable AI technique that quantifies each feature’s contribution to individual predictions, to guide stepwise feature elimination.

The support vector machine maintained the highest AUC throughout the feature reduction process and emerged as the final model, incorporating just eight predictors: coronary artery lesions, C-reactive protein (CRP), age at presentation, neutrophil percentage, white blood cell count, total cholesterol, platelet count, and serum albumin. The eight-feature model outperformed both a 14-feature version (AUC 0.782 versus 0.756) and a minimal two-feature model (0.782 versus 0.736), with differences confirmed by the DeLong test. In internal validation, the final model achieved an AUC of 0.782, and SHAP scatter plots clarified the direction of each predictor’s effect: CRP, neutrophil percentage, white blood cell count, and the presence of coronary artery lesions pushed predictions toward resistance, while older age, higher total cholesterol, higher platelet counts, and higher albumin were protective. For example, white blood cell counts above 21.15 ×10⁹/L and the presence of coronary lesions produced positive SHAP values, whereas patients older than 65 months showed predominantly negative values.

External and prospective validation demonstrated that the model generalizes beyond its original institution. The AUC reached 0.746 in the Fuzhou cohort, 0.759 in Yangzhou, and 0.799 in the prospective Suzhou cohort. Calibration curves showed good agreement between predicted and observed probabilities, with Brier scores of 0.099, 0.097, and 0.068 across the three validation settings. Decision curve analysis, which quantifies net clinical benefit across a range of risk thresholds, showed the model outperformed both the “treat-all” and “treat-none” default strategies across clinically relevant probability ranges in every cohort. When benchmarked against established scoring systems, the machine learning model clearly outperformed the Kobayashi, Egami, Sano, Formosa, and Li scores, whose AUCs ranged from 0.619 to 0.674, all well below the SVM’s 0.782, which also delivered balanced sensitivity of 0.735 and specificity of 0.721.

A key concern in clinical machine learning is the “black box” problem, in which opaque algorithms undermine clinician trust. The researchers addressed this with SHAP-based global and local explanations, showing for individual patients which factors drove their risk classification. Interaction analyses further revealed that albumin had the highest overall interaction strength, followed by CRP, platelets, and total cholesterol, indicating the model captured non-additive relationships among clinical variables rather than simple independent effects. To translate the model into practice, the team built an interactive web-based calculator on the Shiny platform that accepts the eight features and returns a color-coded, individualized risk estimate. The optimal classification threshold, set at 14.9 percent by the Youden index, yielded a sensitivity of 0.648 and specificity of 0.606, and the authors emphasize that the threshold should guide early risk stratification rather than serve as a strict treatment cutoff.

Because coronary artery lesions require echocardiography, which may not be immediately available, the team also reconstructed a model without that predictor. Among algorithms tested without coronary lesions, an SVM variant, LASSO, and elastic net performed comparably, and stepwise reduction identified a five-variable model with an AUC of 0.762 that outperformed both smaller and larger versions. Importantly, the performance difference between the model with coronary lesions (AUC 0.750) and the model without them (AUC 0.739) was not statistically significant, meaning clinicians facing a febrile child before echocardiography can still obtain a meaningful risk estimate from blood tests and age alone.

The biological signals captured by the model align with current understanding of Kawasaki disease pathogenesis. Elevated CRP and neutrophil percentage reflect intense systemic inflammation, and neutrophils are known to contribute to vascular injury by releasing reactive oxygen species, proteases, and proinflammatory cytokines, a mechanism supported by autopsy studies showing dense neutrophilic infiltration in coronary lesions. Thrombocytopenia may indicate ongoing platelet consumption at sites of coronary artery injury, where platelets adhere to exposed collagen and release vascular endothelial growth factor and matrix metalloproteinases that aggravate vascular remodeling. Hypoalbuminemia likely reflects increased vascular permeability and impaired hepatic synthesis under systemic inflammation, while the association between low total cholesterol and resistance is consistent with prior observations of lipid abnormalities in resistant patients. Younger age, an independent risk factor, is thought to relate to immune system immaturity and to atypical or incomplete presentations that delay diagnosis; previous work has reported resistance rates as high as 31.5 percent in infants under 12 months.

The analysis also revealed clear age-related heterogeneity in which biomarkers matter. Among infants aged 0 to 5 months, only neutrophil percentage and total cholesterol differed significantly between resistant and responsive patients, whereas in children aged 3 to 4 years, seven of the eight features showed significant associations. The authors note that modeling age as a single variable may not fully capture such effect modification and suggest that future work could benefit from age-stratified or interaction-based modeling approaches.

The study was reported in accordance with the TRIPOD+AI guidelines for transparent reporting of clinical prediction models using artificial intelligence, and methodological quality was assessed with the PROBAST+AI tool. That assessment identified a high risk of bias in the participants domain owing to the retrospective design of the primary cohort, in the predictors domain because blinding of predictor assessment could not be ensured, and in the analysis domain because sample size considerations were not formally assessed, while the outcome domain was rated low risk and applicability concerns were low across all domains. A conventional logistic regression model built on the same eight predictors achieved an AUC of 0.732, statistically indistinguishable from the SVM’s 0.782, a finding the authors interpret honestly, noting that in settings with moderate sample sizes and structured clinical data, traditional approaches can remain competitive with machine learning.

The authors acknowledge further limitations, including potential clinical subjectivity in Kawasaki disease diagnosis, reliance on structured electronic medical record data without raw echocardiographic images, and possible calibration drift from temporal differences between development and validation cohorts collected over different years. They call for prospective trials using propensity score matching to determine whether model-guided decision-making actually improves clinical outcomes, and they propose future integration of multi-omics and imaging data, as well as exploration of emerging tabular foundation models such as TabPFN. For now, the model and its freely accessible web calculator offer pediatricians a validated, interpretable, and immediately usable instrument for identifying, at the bedside and before the first infusion, which children with Kawasaki disease need escalated therapy to protect their coronary arteries.

Understanding and predicting IVIG resistance remains a high priority because early identification of resistant cases can drive tailored interventions and reduce morbidity. Iwashima’s comprehensive review meticulously charts the trajectory of current predictive models, highlighting both the strides made and the persistent obstacles obstructing their clinical robustness. These models aim to stratify risk by capitalizing on a spectrum of clinical, laboratory, and genetic indicators, yet their practical application has been fraught with multifaceted challenges.

One major evolution in prediction approaches stems from integrating multi-modal data sources. Historically, clinical parameters such as prolonged fever duration, elevated inflammatory markers, and specific hematologic profiles have been cornerstones of resistance prediction. However, these parameters alone often lack sufficient specificity and sensitivity. This gap prompted researchers to incorporate molecular and genomic data, offering richer biological context and potential mechanistic insights.

Genetic markers, particularly polymorphisms associated with immune regulation pathways, have begun to illuminate individual susceptibilities to IVIG resistance. Variations in genes encoding cytokines, immune receptors, and signaling molecules can influence inflammatory responses and consequently the therapeutic efficacy of IVIG. Despite these promising leads, genetic predictors remain limited by population heterogeneity and the complexity of gene-environment interactions, often resulting in inconsistent predictive power across diverse cohorts.

Further complexity arises from the immunopathology of Kawasaki disease itself, which remains only partially understood. The disease elicits a hyperinflammatory state characterized by endotheliitis and dysregulated innate and adaptive immune responses. This heterogeneity in immune activation profiles challenges the development of universally applicable models. Iwashima emphasizes that bridging the gap between immunological mechanisms and predictive algorithms is essential for enhancing model validity.

Machine learning and artificial intelligence (AI) techniques have ushered in a new era of predictive modeling. By harnessing large datasets that include clinical, laboratory, genetic, and even imaging variables, AI models can detect subtle, non-linear patterns inaccessible to traditional statistical analyses. These next-generation models exhibit improved predictive accuracy but require rigorous validation and standardization prior to widespread clinical adoption.

Still, significant hurdles persist in data quality and uniformity. Variability in diagnostic criteria, treatment protocols, and follow-up strategies complicates the aggregation of multi-center data crucial for model training. Moreover, ethical and logistical concerns related to genetic data sharing across regions impede the construction of universally applicable models.

Iwashima’s review also underscores the critical role of biomarker discovery. Emerging data suggest that specific immune mediators, such as N-terminal pro-brain natriuretic peptide (NT-proBNP) and various cytokine profiles, may serve as early indicators of IVIG resistance. Elucidating the temporal dynamics of these biomarkers relative to disease onset and IVIG administration will be pivotal in refining predictive frameworks.

Another dimension addressed is the timing of prediction and intervention. Early prediction is vital to guide intensified therapeutic regimens, including corticosteroids or alternative immunomodulatory agents, which have shown promise in mitigating resistance outcomes. However, premature escalations may expose children to unnecessary risks, underscoring the need for predictive precision.

The author further highlights potential advancements through integrating longitudinal analysis and real-time monitoring technologies. Continuous assessment of clinical and biochemical parameters during hospitalization could enhance dynamic risk stratification. Wearable devices and point-of-care testing might, in the future, feed live data into predictive models, revolutionizing KD management.

Collaboration across disciplines emerges as a key theme. The synergy between pediatricians, immunologists, geneticists, data scientists, and bioinformaticians is imperative to unravel the multifactorial etiology of IVIG resistance and translate this understanding into actionable clinical tools. Cross-disciplinary consortia and large-scale registries will be instrumental in overcoming current limitations.

Despite the challenges, the trajectory outlined by Iwashima is distinctly optimistic. The refinement of IVIG resistance prediction models promises to transform KD care from empiricism toward precision medicine, diminishing adverse cardiovascular sequelae and improving long-term quality of life for affected children. Continuous innovation, coupled with rigorous clinical validation, is essential to realize this vision.

In conclusion, the evolution of IVIG resistance prediction models for Kawasaki disease is emblematic of broader trends in personalized pediatric care. By integrating clinical acumen with cutting-edge computational and molecular approaches, the field edges closer to predictive accuracy that can meaningfully impact therapeutic decisions. While unresolved challenges necessitate ongoing research, the convergence of technology and medical insight signals a paradigm shift in combating resistant Kawasaki disease, heralding a future of safer, more effective patient-centered care.

Subject of Research: Predictive models of intravenous immunoglobulin (IVIG) resistance in Kawasaki disease

Article Title: Evolving strategies and unresolved challenges in IVIG resistance prediction models for Kawasaki disease

Article References:
Iwashima, S. Evolving strategies and unresolved challenges in IVIG resistance prediction models for Kawasaki disease. Pediatr Res (2026). https://doi.org/10.1038/s41390-026-05035-y

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05035-y (05 May 2026)