Development and validation of a novel interpretable machine learning model integrating immune-inflammatory indicators for intravenous immunoglobulin resistance in Kawasaki disease

Abstract

Background

Children with Kawasaki disease (KD) who are resistant to intravenous immunoglobulin (IVIG) therapy face a substantially increased risk of developing coronary artery lesions (CALs). Developing a robust predictive model to identify pediatric patients at high risk of IVIG resistance is crucial for optimizing clinical decision-making and improving prognosis. This study aimed identify risk predictors for IVIG resistance in children with KD and to establish and validate an interpretable machine learning (ML)-based predictive model for clinical application.

Methods

Retrospective analysis was carried out on clinical data sourced from 1,584 KD patients who received initial IVIG treatment during their first hospitalization at Xuzhou Children’s Hospital between January 2019 and December 2024. This cohort was randomly allocated into the training (70%) and test (30%) sets. Six distinct ML algorithms—Light Gradient Boosting Machine (LightGBM), Random Forest, eXtreme Gradient Boosting (XGBoost), Neural Network (NeuralNet), Support Vector Machine (SVM), and ElasticNet Logistic Regression—were employed to develop predictive models. Comparative performance was evaluated employing the area under the receiver operating characteristic curve (AUC). Then, SHapley Additive exPlanations (SHAP) were applied to quantify each variable’s contribution to the optimal model.

Results

The LightGBM model demonstrated superior discriminative performance, attaining an AUC of 0.832 [95% confidence interval (CI): 0.766–0.898] on the independent test set, with a sensitivity of 0.860 and a specificity of 0.639. SHAP summary plots revealed that the five most influential features predicting IVIG resistance were, in descending order: fever duration before initial IVIG, the neutrophil-to-lymphocyte ratio (NLR), interleukin-1β (IL-1β) level, albumin (ALB) level, and aspartate aminotransferase (AST) level.

Conclusions

Our analysis identified five pivotal predictors (fever duration before initial IVIG, NLR, IL-1β, ALB, and AST) for IVIG resistance and validated an interpretable LightGBM model with satisfactory performance. This model shows potential for estimating the risk of IVIG resistance, thereby aiding in the personalized therapeutic strategies for children with KD.

Highlight box.

Key findings

• A novel interpretable Light Gradient Boosting Machine model was developed and validated for predicting intravenous immunoglobulin (IVIG) resistance in Kawasaki disease (KD), with superior predictive performance.

What is known and what is new?

• IVIG resistance elevates coronary artery lesion (CAL) risk in KD; existing prediction models lack consensus standards and show limited generalizability across populations.

• The model incorporated a comprehensive panel of immune-inflammatory markers (neutrophil-to-lymphocyte ratio, T cell subsets, and cytokines) to enhance predictive performance. SHapley Additive Explanations analysis identified interleukin-1β as a key novel predictor unreported in prior prediction models for IVIG resistance in KD.

What is the implication, and what should change now?

• This model enables early identification of high-risk KD patients, facilitating timely intervention to reduce CAL risk; standardization of model construction and routine detection of immune-inflammatory markers are recommended to optimize KD risk stratification.

Introduction

Kawasaki disease (KD) constitutes an acute immune-mediated vasculitis predominantly influencing children under the age of five (1). This condition primarily involves small and medium-sized arteries throughout the body, particularly the coronary arteries, potentially resulting in coronary artery lesions (CALs). KD has emerged as a major cause of acquired heart disease in pediatric populations (2,3). While the disease occurs globally, its incidence is highest in East Asia and demonstrates a gradually increasing trend (2,4). The pathogenesis of KD has not yet been elucidated, and its clinical diagnosis still relies on a comprehensive assessment of typical clinical manifestations and laboratory test results. The standard treatment, intravenous immunoglobulin (IVIG) combined with aspirin, effectively reduces CALs incidence. Nevertheless, roughly 10% to 20% of children do not respond to the initial IVIG infusion, a condition termed IVIG-resistant KD. Although rescue therapies exist—such as a second IVIG infusion, corticosteroids, or infliximab—these patients face a substantially elevated risk of CALs due to persistent vascular inflammation (5-7). This imposes a substantial medical burden on affected children and their families. Accumulating evidence indicates that adjunctive corticosteroid therapy in high-risk IVIG-resistant populations during initial standard treatment can reduce CALs incidence (8-10). Consequently, the early identification of children at high risk for IVIG resistance and the implementation of personalized treatment constitute a critical clinical need for improving long-term outcomes.

Numerous predictive models for IVIG resistance in KD individuals have been developed by researchers worldwide (11-13). Nonetheless, most models rely on traditional logistic regression analyses, incorporate a relatively limited range of clinical and laboratory indicators, and lack standardization. While they exhibit good sensitivity and specificity within their development cohorts, their predictive performance often diminishes in external populations, indicating significant regional and ethnic heterogeneity and limited generalizability. Thus, developing superior predictive models for different regions remains challenging. Xuzhou Children’s Hospital, serving as a regional pediatric medical center for the Huaihai Economic Zone (covering four provinces), caters to a large pediatric population with similar demographic characteristics. This center admits a considerable number of KD patients, reflecting the region’s high KD incidence. Despite this, a risk prediction model for IVIG resistance specific to this population has not yet been established.

Advances in machine learning (ML) technology offer new ways to overcome the limitations of traditional models. ML excels at integrating multidimensional clinical data, optimizing feature selection, and constructing models. ML is effective in clinical domains such as disease diagnosis, outcome prediction, and image interpretation (14-16). SHapley Additive exPlanations (SHAP), a game-theoretic methodology for model interpretation, provides explainability for ML predictions by quantifying each feature’s contribution to the outcome (17).

Accordingly, this large-scale retrospective study incorporated multidimensional data including demographic characteristics, clinical manifestations, routine blood test results, biochemical indices, and immune-inflammatory indicators, and aimed to employ ML algorithms to develop and validate an interpretable prediction model for IVIG-resistant KD. Furthermore, it seeks to identify critical predictors of treatment resistance, thereby providing a clinical decision framework for early intervention to mitigate CALs. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-907/rc).

Methods

Study participants

We retrospectively retrieved electronic medical records of hospitalized patients discharged with the International Classification of Diseases, Tenth Revision (ICD-10) code M30.3 (KD) from the Hospital Information System (HIS) of Xuzhou Children’s Hospital. The present retrospective cohort study incorporated children with a first-time diagnosis of KD admitted to Xuzhou Children’s Hospital between January 2019 and December 2024, all of whom received IVIG. Participants were categorized into IVIG-resistant and IVIG-responsive groups based on treatment efficacy. IVIG resistance was marked by persistent or recrudescence fever lasting ≥36 h following the initial IVIG treatment (5). Inclusion criteria: (I) meeting the American Heart Association (AHA) 2017 diagnostic criteria for complete or incomplete KD (5); (II) aged ≤18 years; (III) having comprehensive clinical data and laboratory indicators available prior to the first dose of IVIG. Exclusion criteria (any of the following): (I) fever persisting for more than 10 days prior to initial IVIG treatment; (II) administration of corticosteroids or other immunosuppressants during or before the initial IVIG treatment; (III) prior IVIG treatment at another hospital; (IV) concurrent severe infection or immunodeficiency disease; (V) history of recurrent KD; (VI) were IVIG-naive. The study period coincided with the peak incidence of multisystem inflammatory syndrome in children (MIS-C). Between 2020 and 2022, stringent non-pharmaceutical interventions (NPIs) implemented in China resulted in no detected SARS-CoV-2 infections among hospitalized children in our cohort, and no cases met MIS-C diagnostic criteria (18,19). Following policy relaxation in 2023, cases diagnosed with MIS-C according to the 2022 surveillance case definition from the Centers for Disease Control and Prevention (CDC) were not included in this study (20). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Xuzhou Children’s Hospital (No. 2024-05-17-H17) and informed consent was obtained from all patients’ parents or legal guardians.

Information collection

Trained personnel reviewed electronic medical records to collect demographic data, clinical characteristics, and laboratory results obtained before the first IVIG treatment. Demographic information included age and sex. Clinical characteristics comprised fever duration before initial IVIG and IVIG treatment response. Laboratory investigations encompassed various hematological, biochemical, inflammatory, and immunological parameters: white blood cell count (WBC), platelet count (PLT), hemoglobin (HB), albumin (ALB), alanine aminotransferase (ALT), aspartate aminotransferase (AST), serum sodium (Na), direct bilirubin (DBil), gamma-glutamyl transferase (GGT), erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), procalcitonin (PCT), CD3⁺ T-lymphocyte count (CD3⁺), CD4⁺ T-lymphocyte count (CD4⁺), CD8⁺ T-lymphocyte count (CD8⁺), interleukin-1β (IL-1β), interleukin-6 (IL-6), interleukin-10 (IL-10), interferon-γ (IFN-γ), and tumor necrosis factor-α (TNF-α). The neutrophil-to-lymphocyte ratio (NLR) was derived from laboratory data. Feature selection was based on clinical experience, relevant literature, and clinical availability. The detection of cytokines and T lymphocyte subsets was performed in accordance with the standardized diagnosis and treatment protocol for KD established at our institution. Specifically, cytokines were quantified via multiplex immunoassay based on flow cytometry, reducing the total assay turnaround time to 4 to 6 hours (21). This methodology fully complies with the requirements of the clinical laboratory quality management system. For multiple pre-treatment laboratory tests, the most aberrant result was used for analysis.

Machine learning model (MLM) development

Data preparation and preprocessing

The original dataset comprised 24 variables, excluding the outcome variable. Both sex and the outcome variable were defined as binary. Only continuous variables exhibited missing values, with all missing rates below 5%. Missing data were imputed using the missForest algorithm, a non-parametric, iterative random forest-based method that simultaneously handles continuous and categorical variables, captures nonlinear relationships and interactions, and does not require distributional assumptions (22). To mitigate the possible influence of highly skewed distributions on model performance and achieve more stable variance, a log(1+x) transformation was applied to all numerical features of the independent variables. To evaluate and address multicollinearity, variance inflation factor (VIF) analysis with a threshold of five was adopted for stepwise feature selection, leading to the exclusion of three highly collinear variables (Figure S1). Consequently, 21 variables were retained for subsequent analyses. This dataset was randomly split into the training (70%) and test (30%) sets. To handle class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training set to generate synthetic instances for the minority class, thereby balancing the class distribution for improved model learning (23).

Model development and hyperparameter optimization

Six ML algorithms—Light Gradient Boosting Machine (LightGBM), Random Forest, eXtreme Gradient Boosting (XGBoost), Neural Network (NeuralNet), Support Vector Machine (SVM), and ElasticNet Logistic Regression—were employed to create prediction models using the training set. Bayesian optimization was utilized for hyperparameter tuning for all models. A consistent 5-fold cross-validation strategy was deployed on the training set to evaluate performance. After identifying the optimal hyperparameter configuration, the final model was trained on the entire training set using these parameters.

Model performance evaluation

Model performance was evaluated on the independent test set. Performance metrics, including sensitivity, specificity, accuracy, and the area under the receiver operating characteristic (ROC) curve (AUC), were computed utilizing the confusion matrix, defining class “1” as the positive class. ROC curves for all models on the test set were plotted. The AUCs, along with 95% confidence intervals (CIs), were reported to facilitate comparison of classification performance.

Model interpretation

To elucidate the prediction mechanism of the optimal model, SHAP analysis was carried out. SHAP values were computed for each feature to quantify its contribution to the model’s predictions, thereby revealing feature importance and directional impact. Higher mean absolute SHAP values indicate greater feature importance, while lower values suggest lesser influence. SHAP values can be either positive or negative, with positive values indicating a positive influence on the prediction and an increase in the predicted value, while negative values have the opposite effect. The visualization of SHAP analysis includes feature importance bar charts, feature importance beeswarm plots, individual sample force plots, and feature dependence plots to comprehensively explain the model’s decision-making process.

Statistical analysis

Baseline characteristics were analyzed using SPSS 27.0. Normality was assessed for continuous variables. Normally distributed data were reported as mean ± standard deviation (SD) and compared using the independent-samples t-test; non-normally distributed data were summarized as median [interquartile range (IQR)] and compared using the Mann-Whitney U test. Categorical data were expressed as counts (proportions) and compared using the Chi-squared test. Statistical significance was defined as P<0.05.

All ML procedures—including model development, validation, and evaluation—were implemented in R (version 4.3.3). Data preprocessing was performed utilizing the ‘tidyverse’ package; missing data imputation was performed using the ‘missForest’ package; data splitting and confusion matrix calculations were carried out with the ‘caret’ package; class imbalance was addressed via SMOTE using the ‘smotefamily’ package; VIFs were computed using the ‘car’ package. The six MLMs were implemented using the following R packages: ‘randomForest’ (Random Forest), ‘lightgbm’ (LightGBM), ‘e1071’ (SVM), ‘nnet’ (NeuralNet), ‘xgboost’ (XGBoost), and ‘glmnet’ (ElasticNet). Hyperparameter tuning was achieved through Bayesian optimization using the ‘ParBayesianOptimization’ package. Interpretability of the models was assessed using SHAP, with analyses and visualizations generated via the ‘fastshap’ and ‘shapviz’ packages.

Results

Baseline clinical characteristics

Among 1,730 eligible KD patients (Figure 1), 1,584 were included after applying exclusion criteria. The median age was 22 months (IQR, 11–40 months), and 976 (61.62%) were male. The cohort incorporated 144 (9.09%) IVIG-resistant and 1,440 IVIG-responsive patients. Relative to the responsive group, the resistant group exhibited notably elevated levels of WBC, NLR, ALT, AST, DBil, GGT, CRP, PCT, IL-10, IL-6, IL-1β, IFN-γ, and TNF-α, alongside markedly lower levels of PLT, ALB, Na, CD3⁺, CD4⁺, CD8⁺, and a shorter fever duration before initial IVIG (P<0.05). The two groups exhibited no significant differences in age, ESR, or HB levels (P>0.05; Table 1).

Figure 1.

Study flowchart of patient selection and model development. IVIG, intravenous immunoglobulin.

Table 1. Comparison of demographic, clinical, and laboratory characteristics between IVIG-resistant and IVIG-responsive KD patients.

Variables	IVIG-responsive (n=1,440)	IVIG-resistance (n=144)	Statistic	P
Age, months	22.00 (11.00, 40.00)	21.00 (12.00, 37.25)	Z=−0.21	0.84
Sex			χ2=6.58	0.01
Male	873 (60.62)	103 (71.53)
Female	567 (39.38)	41 (28.47)
Fever before IVIG, d	6.00 (5.00, 7.00)	5.00 (5.00, 6.00)	Z=−6.77	<0.001
CRP, mg∙L−1	64.66 (38.19, 109.61)	89.47 (53.27, 131.78)	Z=−4.45	<0.001
ESR, mm∙h−1	52.00 (37.00, 65.43)	50.22 (37.00, 64.00)	Z=−0.32	0.75
WBC, ×109∙L−1	14.97 (11.74, 18.49)	16.95 (12.99, 21.30)	Z=−3.43	<0.001
NLR	2.62 (1.58, 4.72)	5.22 (3.11, 9.63)	Z=−8.30	<0.001
PLT, ×109∙L−1	354.00 (276.75, 435.00)	306.50 (232.50, 385.00)	Z=−4.04	<0.001
HB, g∙L−1	111.00 (104.00, 117.00)	110.50 (102.00, 118.00)	Z=−0.76	0.45
PCT, ng∙mL−1	0.32 (0.13, 0.76)	0.93 (0.32, 2.21)	Z=−6.72	<0.001
ALB, g∙L−1	38.50 (36.00, 40.70)	36.85 (33.68, 40.30)	Z=−3.34	<0.001
ALT, U∙L−1	24.00 (14.00, 66.00)	52.50 (20.00, 208.00)	Z=−5.61	<0.001
AST, U∙L−1	31.00 (24.00, 46.00)	44.00 (29.00, 114.75)	Z=−5.67	<0.001
Na, mmol∙L−1	136.90 (135.00, 139.00)	135.60 (133.95, 137.53)	Z=−4.75	<0.001
DBil, μmol∙L−1	2.40 (1.70, 3.70)	3.85 (2.18, 7.35)	Z=−6.36	<0.001
GGT, U∙L−1	23.00 (12.64, 84.25)	61.00 (18.00, 167.00)	Z=−4.68	<0.001
CD3+, cells∙μL−1	2,497.00 (1,582.50, 3,607.43)	1,509.47 (912.84, 2,638.31)	Z=−6.86	<0.001
CD4+, cells∙μL−1	1,471.55 (862.75, 2,156.99)	852.88 (461.30, 1,534.09)	Z=−6.99	<0.001
CD8+, cells∙μL−1	864.97 (551.76, 1,257.81)	534.50 (303.98, 895.08)	Z=−6.53	<0.001
IL-1β, pg∙mL−1	8.50 (2.90, 17.42)	12.04 (3.00, 39.23)	Z=−2.65	0.008
IL-6, pg∙mL−1	43.76 (17.60, 105.42)	101.05 (25.30, 254.55)	Z=−4.96	<0.001
IL-10, pg∙mL−1	9.40 (3.68, 19.90)	23.65 (7.40, 62.05)	Z=−6.86	<0.001
IFN-γ, pg∙mL−1	8.20 (2.70, 15.70)	14.23 (4.90, 30.34)	Z=−5.01	<0.001
TNF-α, pg∙mL−1	2.70 (1.20, 4.44)	3.47 (1.30, 6.35)	Z=−2.66	0.008

Data are presented as M (Q1, Q3) or n (%). ALB, albumin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; CRP, C-reactive protein; DBil, direct bilirubin; ESR, erythrocyte sedimentation rate; GGT, gamma-glutamyl transferase; HB, hemoglobin; IFN-γ, interferon-gamma; IL-10, interleukin-10; IL-1β, interleukin-1β; IL-6, interleukin-6; IVIG, intravenous immunoglobulin; KD, Kawasaki disease; Na, serum sodium; NLR, neutrophil-to-lymphocyte ratio; PCT, procalcitonin; PLT, platelet count; TNF-α, tumor necrosis factor-alpha; WBC, white blood cell count.

Baseline characteristics of test and training sets

The dataset was randomly allocated into the test (n=475, 30%) and training (n=1,109, 70%) sets. Comparison of baseline characteristics across the two sets revealed a statistically significant difference only in IFN-γ levels (P<0.05). Other indicators showed no significant differences (P>0.05; Table 2).

Table 2. Comparison of baseline characteristics across the training and test sets.

Variables	Test set (n=475)	Training set (n=1,109)	Statistic	P
Age, months	24.00 (12.00, 43.00)	21.00 (11.00, 38.00)	Z=−1.73	0.08
Sex			χ2=0.68	0.41
Male	300 (63.16)	676 (60.96)
Female	175 (36.84)	433 (39.04)
Fever before IVIG, d	6.00 (5.00, 7.00)	6.00 (5.00, 7.00)	Z=−0.90	0.37
Outcome			χ2=0.00	0.97
IVIG-responsive	432 (90.95)	1008 (90.89)
IVIG-resistant	43 (9.05)	101 (9.11)
CRP, mg∙L−1	65.60 (38.94, 113.79)	66.33 (39.29, 110.70)	Z=−0.03	0.97
ESR, mm∙h−1	52.66 (39.00, 66.00)	52.00 (37.00, 65.00)	Z=−0.95	0.34
WBC, ×109∙L−1	15.20 (11.85, 18.73)	14.99 (11.81, 18.83)	Z=−0.15	0.88
NLR	2.76 (1.67, 5.33)	2.80 (1.69, 5.09)	Z=−0.39	0.70
PLT, ×109∙L−1	344.00 (273.50, 429.50)	353.00 (272.00, 435.00)	Z=−0.69	0.49
HB, g∙L−1	111.00 (104.00, 118.00)	110.00 (103.00, 117.00)	Z=−1.34	0.18
PCT, ng∙mL−1	0.35 (0.13, 0.82)	0.34 (0.13, 0.91)	Z=−0.20	0.84
ALB, g∙L−1	38.10 (35.30, 40.45)	38.50 (36.00, 40.70)	Z=−1.67	0.10
AST, U∙L−1	32.00 (25.00, 49.00)	32.00 (24.00, 48.00)	Z=−0.56	0.58
Na, mmol∙L−1	136.30 (134.95, 139.00)	136.84 (135.00, 139.00)	Z=−1.48	0.14
DBil, μmol∙L−1	2.50 (1.70, 3.80)	2.50 (1.70, 4.00)	Z=−0.52	0.60
GGT, U∙L−1	23.00 (13.00, 97.50)	25.00 (13.00, 90.00)	Z=−0.22	0.83
CD8+, cells∙μL−1	844.00 (522.00, 1,185.00)	853.16 (528.00, 1,233.65)	Z=−0.61	0.54
IL-1β, pg∙mL−1	8.47 (2.70, 16.21)	8.53 (3.00, 19.20)	Z=−1.63	0.10
IL-6, pg∙mL−1	45.88 (17.80, 107.21)	46.81 (18.94, 120.70)	Z=−0.48	0.63
IL-10, pg∙mL−1	10.20 (3.90, 21.33)	9.90 (3.80, 23.30)	Z=−0.06	0.96
IFN-γ, pg∙mL−1	7.90 (2.50, 14.90)	8.90 (3.00, 17.40)	Z=−2.01	0.044
TNF-α, pg∙mL−1	2.57 (1.15, 4.36)	2.80 (1.30, 4.70)	Z=−1.72	0.09

Data are presented as M (Q1, Q3) or n (%). ALB, albumin; AST, aspartate aminotransferase; CRP, C-reactive protein; DBil, direct bilirubin; ESR, erythrocyte sedimentation rate; GGT, gamma-glutamyl transferase; HB, hemoglobin; IFN-γ, interferon-gamma; IL-10, interleukin-10; IL-1β, interleukin-1β; IL-6, interleukin-6; IVIG, intravenous immunoglobulin; Na, serum sodium; NLR, neutrophil-to-lymphocyte ratio; PCT, procalcitonin; PLT, platelet count; TNF-α, tumor necrosis factor-alpha; WBC, white blood cell count.

MLM performance

The ROC curves for the six MLMs on the test set are presented in Figure 2. The LightGBM model attained the highest AUC of 0.832 (95% CI: 0.766–0.898) on the test set, indicating superior predictive performance among the evaluated models. This was followed by the Random Forest model (AUC =0.814, 95% CI: 0.749–0.879). The AUC values for the remaining models were: ElasticNet (0.796, 95% CI: 0.720–0.872), XGBoost (0.792, 95% CI: 0.721–0.863), SVM (0.714, 95% CI: 0.632–0.796), and NeuralNet (0.681, 95% CI: 0.588–0.773). The LightGBM model achieved a sensitivity of 0.860 and a specificity of 0.639 (Table 3).

Figure 2.

ROC curves of the six MLMs on the test set. AUC, area under the receiver operating characteristic curve; CI, confidence interval; LightGBM, Light Gradient Boosting Machine; MLM, machine learning model; NeuralNet, Neural Network; ROC, receiver operating characteristic; SVM, Support Vector Machine; XGBoost, eXtreme Gradient Boosting.

Table 3. Predictive performance of the six models on the test set.

Models	AUC (95% CI)	Sensitivity	Specificity	Accuracy
LightGBM	0.832 (0.766–0.898)	0.860	0.639	0.659
RandomForest	0.814 (0.749–0.879)	0.721	0.792	0.785
ElasticNet	0.796 (0.720–0.872)	0.721	0.752	0.749
XGBoost	0.792 (0.721–0.863)	0.767	0.720	0.724
SVM	0.714 (0.632–0.796)	0.233	0.917	0.855
NeuralNet	0.681 (0.588–0.773)	0.349	0.929	0.876

AUC, area under the receiver operating characteristic curve; CI, confidence interval; LightGBM, Light Gradient Boosting Machine; NeuralNet, Neural Network; SVM, Support Vector Machine; XGBoost, eXtreme Gradient Boosting.

SHAP interpretability analysis

SHAP analysis was implemented on the LightGBM model to interpret feature importance. Figure 3A shows that the top five features most influential in predicting IVIG resistance were: fever duration before initial IVIG (mean |SHAP| value =0.043), NLR (0.030), IL-1β (0.023), ALB (0.022), and AST (0.022). The risk of IVIG resistance increased with higher NLR, IL-1β, and AST levels, and with shorter fever duration before initial IVIG and lower ALB levels (Figure 3B,3C and Figure 4). Notably, sex ranked sixth, indicating a contribution to predicting IVIG resistance.

Figure 3.

SHAP analysis for model interpretation. (A) SHAP feature importance bar chart: ranking of feature importance in the LightGBM model. (B) SHAP feature importance beeswarm plot: impact of different feature value ranges on model predictions. Purple indicates lower feature values, yellow denotes higher values. (C) SHAP force plot for an individual sample: showing how each feature shifts the model’s output from the base value. Red arrows represent features decreasing the prediction, yellow arrows signify features increasing it. ALB, albumin; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; IL-10, interleukin-10; IL-1β, interleukin-1β; IL-6, interleukin-6; IVIG, intravenous immunoglobulin; LightGBM, Light Gradient Boosting Machine; Na, serum sodium; NLR, neutrophil-to-lymphocyte ratio; PCT, procalcitonin; PLT, platelet count; SHAP, SHapley Additive exPlanations; TNF-α, tumor necrosis factor-α; WBC, white blood cell count.

Figure 4.

SHAP feature dependence plots of top nine features: visualize the correlation of individual feature values with their SHAP values, revealing the feature’s impact pattern on model predictions. ALB, albumin; AST, aspartate aminotransferase; IL-10, interleukin-10; IL-1β, interleukin-1β; IVIG, intravenous immunoglobulin; NLR, neutrophil-to-lymphocyte ratio; SHAP, SHapley Additive exPlanations; WBC, white blood cell count.

Sex subgroup analysis

To explore sex-specific predictors of IVIG resistance, separate LightGBM models were developed for males (n=976) and females (n=608). SHAP analysis (Figure 5) illustrated variations in the importance of predictive variables between the two groups. For males, the five most significant features included fever duration before initial IVIG, NLR, ALB, age, and IFN-γ (Figure 5A,5C). In females, the top five contributors were NLR, IL-1β, ESR, GGT, and AST (Figure 5B,5D). NLR remained a highly influential predictor across all cohorts.

Figure 5.

SHAP feature importance plots for sex subgroups. (A) Feature importance bar chart: male subgroup. (B) Feature importance bar chart: female subgroup. (C) SHAP beeswarm plot: male subgroup. (D) SHAP beeswarm plot: female subgroup. ALB, albumin; AST, aspartate aminotransferase; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; GGT, gamma-glutamyl transferase; HB, hemoglobin; IFN-γ, interferon-γ; IL-10, interleukin-10; IL-1β, interleukin-1β; IL-6, interleukin-6; IVIG, intravenous immunoglobulin; Na, serum sodium; NLR, neutrophil-to-lymphocyte ratio; PCT, procalcitonin; PLT, platelet count; SHAP, SHapley Additive exPlanations; TNF-α, tumor necrosis factor-α; WBC, white blood cell count.

Discussion

CALs are the most prevalent and serious complication of KD, potentially leading to coronary aneurysm, myocardial infarction, or sudden death (24). A meta-analysis confirmed IVIG resistance as an autonomous predictor of CALs in KD patients (25). Consequently, early prediction has become a focal point of clinical research. This study leveraged large-scale retrospective data and multiple MLMs to predict IVIG resistance, identifying key predictors. The objective was to assist clinicians in early identification of high-risk IVIG-resistant children, offering new insights for risk assessment and individualized treatment.

Over the past two decades, researchers worldwide have devoted their efforts to developing prediction models for IVIG-resistant KD, with internationally recognized examples including the Japanese Kobayashi (26) and Egami (27) scores. However, subsequent validations in European, American, and Chinese populations revealed that while these scores maintain moderate specificity, their sensitivity is generally low, insufficient for effective identification of IVIG non-responders (11,28-32). Concurrently, Chinese investigators have developed prediction models for populations in Beijing, Shanghai, Suzhou, Chongqing, and Taiwan (32-36). Nevertheless, these models vary considerably in their scoring items. Their predictive performance (sensitivity and specificity) is inconsistent and has failed to establish a widely validated and accepted consensus standard. This lack of generalizability is due to regional and ethnic variations affecting model performance across different populations. It is also due to the inherent limitations of the primary statistical method used, traditional logistic regression, in handling clinical data with multiple variables and potential complex nonlinear relationships.

The rapid advancement of artificial intelligence, particularly ML, offers significant advantages for disease prediction, especially when analyzing large, high-dimensional datasets (37,38). Several studies have applied ML to predict IVIG resistance. However, as Mirata et al. (39) noted, these studies demonstrate substantial heterogeneity in model design and parameter selection. They also commonly face challenges related to limited sample size—approximately two-thirds had samples <1,000—and inadequate handling of class imbalance. Against this backdrop, our study retrospectively analyzed 1,584 KD patients and systematically compared six MLMs in terms of predictive performance. To address class imbalance, SMOTE was applied to the training dataset. The validation results indicated that the LightGBM model performed the best overall. It achieved an AUC of 0.832 (95% CI: 0.766–0.898) and a sensitivity exceeding 80%. These results indicate the model’s robust capacity to detect positive instances while maintaining high discriminatory power, which is clinically significant for early warning. As a gradient-boosting decision tree algorithm, LightGBM efficiently processes high-dimensional, sparse medical features, thereby enhancing training speed and resource efficiency. This makes it feasible for large cohort studies. Furthermore, LightGBM’s gradient-based one-sided sampling technique automatically focuses on informative minority class samples, improving the detection rate for high-risk individuals (40,41).

Feature selection is crucial in building multidimensional prediction models. Identifying the most explanatory variables from vast clinical data remains challenging. This study employed SHAP to quantitatively analyze each feature’s contribution, significantly enhancing model interpretability. The analysis based on SHAP values from the LightGBM model identified fever duration before initial IVIG, NLR, IL-1β, ALB, and AST as key predictors.

While the 2017 AHA guidelines (5) recommend that children with KD should start IVIG therapy as soon as possible within 10 days of onset, the 2020 diagnostic criteria for KD by the Japanese Pediatric Society (42) no longer emphasize the duration of fever. Our findings demonstrated that fever duration before initial IVIG dose was a dominant predictor. This aligns with observations reported by Kobayashi et al. (26), Egami et al. (27), and Fu et al. (33). Current evidence indicates that children who are diagnosed and treated early exhibit more intense inflammatory responses, potentially elevating risks of IVIG resistance and CALs. This warrants consideration of intensified therapeutic regimens.

The precise etiology of KD remains elusive, but the prevailing hypothesis is that genetically susceptible individuals experience an unknown antigenic trigger that activates an immune-mediated inflammatory cascade leading to systemic vasculitis (43-45). This process involves complex interactions between innate and adaptive immunity. Immune cell dysregulation and abnormal cytokine secretion characterize KD’s inflammatory profile, underpinning the therapeutic rationale for IVIG (46,47). Therefore, our model innovatively incorporated a comprehensive panel of immune-inflammatory markers, including NLR, T cell subsets, and multiple cytokines—a strategy less prevalent in prior studies. NLR was the second most important predictor, highlighting its significant contribution. As a composite index, NLR may more accurately reflect the imbalance between inflammatory activation and immune regulation in KD than isolated neutrophil percentage or lymphocyte counts. This likely involves neutrophil hyperactivation coupled with lymphocytopenia and functional impairment of lymphocytes, consistent with the conclusions drawn by Kawamura et al. (48) and Li et al. (49).

SHAP-based interpretability analysis further revealed that IL-1β was among the top three with the highest predictive contribution. As a key pro-inflammatory cytokine, IL-1β is extensively implicated in studies of inflammation-related diseases. Gene and transcriptome analyses of blood samples from children with KD, along with experimental evidence from mouse models, confirm that IL-1β plays a critical role in KD pathogenesis. Specifically, NLRP3 inflammasome-mediated maturation and release of IL-1β drive coronary vasculitis development and progression in KD (46,50). Furthermore, IL-1β gene polymorphisms exacerbate inflammatory response by modulating cytokine expression levels and demonstrate a significant association with IVIG resistance (51). Notably, the IL-1 receptor antagonist anakinra has achieved therapeutic success in IVIG-resistant KD patients through targeted blockade of IL-1β signaling (52). Anakinra has been listed as an additional treatment for refractory KD in the 2024 AHA Scientific Statement on the Diagnosis and Treatment of KD (1). These findings collectively validate the centrality of NLR and IL-1β in predicting treatment resistance while providing quantitative evidence of their immunomodulatory mechanisms.

This study validated hypoalbuminemia (ALB) and elevated AST as significant IVIG resistance predictors, aligning with their roles in established clinical scores (32,35,36). Pathophysiologically, reduced ALB may result from increased vascular permeability and suppressed hepatic synthesis due to systemic inflammation. Elevated AST indicates hepatocyte damage from liver involvement. Both abnormalities signify a more severe inflammatory state during KD acute phase.

Given the significant influence of sex on treatment outcomes, we performed a stratified analysis by sex. The results revealed that NLR maintained robust predictive capability for IVIG resistance across both male and female subgroups. These findings support the hypothesis that IVIG resistance reflects a dysregulated systemic inflammatory state. For male children, particular emphasis should be placed on the timing of the initial IVIG infusion and age at disease onset. Additionally, closer post-IVIG infusion monitoring is recommended to facilitate early intervention in cases of therapeutic failure and reduce the risk of CALs. For female children, beyond routine inflammatory markers, enhanced monitoring of liver-related parameters (e.g., GGT and AST) is warranted. These parameters may serve as early indicators of treatment response.

The data-driven feature selection (e.g., based on SHAP values) minimizes bias from prior assumptions, delivering clinicians with an intuitive and reliable ranking of feature importance to guide diagnostic and therapeutic decision-making.

This study extends beyond validating the predictive value of well-documented risk factors to develop a multidimensional predictor panel that preserves the clinical convenience and cost-effectiveness of conventional indicators. Integrating immune-inflammatory indicators such as IL-1β further enhances the model’s ability to capture immune dysregulation, the central pathophysiological mechanism underlying KD. Notably, as a key predictive factor, IL-1β is quantified via a flow cytometry-based multiplex immunoassay, a technique increasingly adopted in tertiary pediatric centers. This method reduces the assay turnaround time to 4 to 6 hours and, given that children with KD who require IVIG treatment need hospitalization, enables routine measurement in well-resourced pediatric inpatient settings.

However, several limitations warrant consideration. First, the single-center, retrospective design may limit generalizability, particularly given regional variations in clinical practices. Moreover, the immune-inflammatory indicators incorporated in our model are not routinely available in primary or most secondary hospitals, which could hinder external validation and broader clinical adoption. Second, prospective validation is required to fully assess model performance, as it has not yet been integrated into clinical workflows, leaving its real-world utility unverified.

To address these challenges, prospective, multicenter, large-scale cohort studies are essential to confirm the model’s applicability through rigorous external validation. We propose a phased validation strategy, which will be initially deployed in tertiary pediatric centers with the requisite laboratory infrastructure. Subsequently, by leveraging regional medical centers as referral hubs, the model can be applied to the management of patients referred from lower-tier institutions, thereby expanding its clinical reach. Additionally, the optimized model will be embedded into the HIS or developed as a dedicated clinical decision support tool, transforming it from a research prototype into an actionable instrument to enable individualized KD management in routine clinical practice.

Conclusions

The present study established a novel ML-based predictive model for early identification of IVIG resistance in KD, with fever duration before initial IVIG, NLR, IL-1β, ALB, and AST identified as critical predictors. By enabling early risk stratification, this model provides a clinical basis for implementing individualized therapies to mitigate CALs in high-risk populations.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Xuzhou Children’s Hospital (No. 2024-05-17-H17) and informed consent was obtained from all patients’ parents or legal guardians.

Data Sharing Statement

Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-907/dss

tp-15-03-74-dss.pdf ^{(88.9KB, pdf)}

DOI: 10.21037/tp-2025-1-907