Kawasaki disease (KD) is an acute vasculitis that primarily affects children under five years old (1). KD primarily manifests as non-exudative bilateral conjunctival hyperemia, oral mucosal changes, non-purulent cervical lymphadenopathy, polymorphic rash, and acute-phase erythema or edema of the hands and feet, or subacute-phase periungual desquamation. Coronary artery lesions (CALs) are the most frequent complication of KD and the leading cause of acquired heart disease in children (2, 3). High-dose intravenous immunoglobulin (IVIG; 2 g/kg) significantly reduces the risk of KD and CALs (4, 5). However, 10–20% of patients do not respond to IVIG, substantially increasing the risk of developing CALs (6, 7). Moreover, more than one IVIG treatment places economic strain on the family. Thus, a tool to accurately predict IVIG non-responders could help develop early intervention strategies, implement more effective alternative treatments for high-risk patients, improve prognosis, and reduce economic pressures. Some scoring models for predicting IVIG resistance exist (8–10), but their practical application is limited, particularly in Northeast China. To date, prediction models for Chinese patients have been developed using regional populations, not national populations, and have not been externally validated.

Machine learning algorithms are advantageous in the medical field given their efficient data processing capabilities, excellent pattern recognition and prediction functions, and ability to handle complex non-linear relationships (11). More specifically, the pattern recognition and prediction performance of machine learning algorithms are remarkable, as they can identify hidden patterns in complex datasets and predict the likelihood of disease development and treatment effects, which has valuable implications for early disease prevention and the development of personalized treatment plans. Medical data are often multidimensional and involve complex non-linear relationships, and machine learning algorithms are particularly well-suited for handling this type of data because they can uncover complex interactions and effects among variables. However, this is difficult to achieve using traditional statistical methods.

Few machine learning applications have been developed for KD, and none are available on a national scale (12). Therefore, this study used advanced machine learning models to predict cases of IVIG resistance in KD.

Machine-learning algorithms are widely utilized in the diagnosis and prognosis of cardiovascular and cerebrovascular disorders, diabetes, and other conditions (16–18). This study used machine learning algorithms to develop a clinical prediction model for IVIG unresponsiveness in patients with KD for the first time. The CatBoost model, based on a large dataset (n = 4,704) from patients in Northeast China, had the best performance of ten machine learning algorithms. Internal validation of the model using 10-fold cross-validation confirmed its excellent performance and stability, and the external validation results reflect the model’s generalizability and robustness, making the prediction model applicable to the whole country.

This study provides solid support for assessing the risk factors for IVIG unresponsiveness in patients with KD in China. Our evaluation of ten machine learning models resulted in a reliable prediction tool for clinical use that, unlike prior KD prediction models (8–10), was interpretable and could visualize each variable’s impact on the prediction findings. To build the model, key characteristics were screened using LASSO regression to eliminate confounding factors, followed by univariable and multivariable logistic regression analyses. To further understand the fundamental aspects of IVIG resistance, we used the SHAP algorithm to analyze the CatBoost model, which identified eight features with a considerable impact on IVIG resistance: winter onset, chlamydial infections, conjunctival congestion, fever duration, and hemoglobin, NT-proBNP, serum ALB, and TBIL levels. The top three variables influencing the model’s predictions were serum TBIL and ALB levels and fever duration.

CatBoost is a gradient-boosting algorithm for categorical data and high-dimensional features (19). Its intrinsic properties make it one of the best choices for real-world applications dealing with categorical variables efficiently while avoiding time-consuming data preprocessing (20). CatBoost significantly suppresses overfitting and enhances the model’s generalization capabilities by employing a novel stochastic alignment technique, particularly when data are restricted. Compared with other gradient boosting algorithms (for example, XGBoost), CatBoost is more efficient and less demanding for hyperparameter adjustment, trains more quickly, and creates superior models (21). The AUCs for the training and validation sets for our CatBoost model were 0·960 and 0·862, respectively, indicating minimal overfitting and good generalizability. The accuracy, sensitivity, and specificity of the CatBoost model were significantly higher than those of other models, making it more suited for the balanced detection of IVIG-sensitive and IVIG-unresponsive events. Finally, we externally verified our CatBoost risk prediction model using data from 568 prospectively enrolled patients with KD at Guangzhou Women and Children’s Medical Center. The sensitivity, specificity, accuracy, and AUC values were 0·817, 0·838, 0·835, and 0·834 (95% CI: 0·783–0·884), respectively, confirming the model’s generalizability and robustness.

Interestingly, this study discovered for the first time that winter and chlamydial infections significantly enhanced the probability of IVIG resistance, validating the idea that environmental variables affect the IVIG response (22). Winter climate and environmental changes may significantly influence immunological functions, contributing to the development of IVIG resistance in patients with KD. Although winter has previously been identified as the peak season for KD (23), our results imply that an elevated risk of winter onset may be linked to the incidence of winter respiratory infections. Respiratory infection pathogens (e.g. influenza virus and adenovirus) in winter cause an increased inflammatory response, increasing the likelihood of IVIG resistance (24). Chlamydial infections are also more frequent in winter, and the systemic inflammatory response they cause affects the course of KD by stimulating the host’s immune system (25–27). Chlamydial infection induces the release of pro-inflammatory cytokines, such as interleukin (IL)-6 and tumour necrosis factor-alpha (TNF-α), and activates the innate immune response, which subsequently amplifies inflammation, particularly in patients with IVIG-unresponsive KD (28). Chlamydial infection induces a Th1-type immune response, decreasing the anti-inflammatory effects of IVIG (29). Furthermore, immunological changes, such as an active complement system and enhanced macrophage responses during winter, may impair the effectiveness of IVIG (26, 27). Such immune system modifications aggravate the pathogenic potential of KD and reduce the treatment efficacy. Infection-driven hyperinflammatory responses activate multiple immune signaling pathways, including TNF-α and IL-1, which are pivotal in the pathogenesis of IVIG-resistant KD. Infection-induced cytokine release stimulates the nuclear factor kappa-light-chain-enhancer of activated B-cells (NF-κB) pathway, which induces a pro-inflammatory response and causes vascular endothelial damage and coronary artery disease. The NF-κB signaling pathway is related to winter-onset viral infections (30). Therefore, identifying winter onset and chlamydial infection as risk factors for IVIG non-responsiveness provides crucial guidelines for treating children with KD.

We also identified elevated TBIL levels as a risk factor for IVIG resistance in patients with KD. High bilirubin levels may indicate a more severe hepatic inflammatory burden, suggesting disease aggravation (31). Sunaga et al. (32) found that a low TBIL level before IVIG therapy indicated a poor immunological response, which might be related to IVIG resistance. Low bilirubin levels indicate insufficient inflammatory pathway activation, which may result in poor IVIG effectiveness. Children in the Tohoku area may be more prone to cholestasis and hyperbilirubinemia. However, Japanese individuals with Sunaga Y disease exhibit different pathological characteristics. Bilirubin has immunomodulatory properties and acts as an anti-inflammatory agent by blocking T-cell activation and decreasing oxidative stress (33). Low bilirubin levels may be ineffective in reducing excessive inflammation, resulting in IVIG resistance (34). In contrast, elevated bilirubin levels may indicate serious hepatic damage, impairing bile excretion due to increased inflammation and increasing the risk of IVIG therapy failure (35). This biphasic effect may explain the conflicting presentation of TBIL levels and IVIG resistance in several studies.

This study has one key limitation. The development phase data and the external validation data in this study were all collected retrospectively, potentially limiting the ability to control for confounding factors. Future studies should address this issue by including more diverse datasets and using prospective study designs. And this study applied the 2017 AHA diagnostic criteria for consistency. Future prospective validation will include patients diagnosed under the latest standards to further confirm the model’s generalizability.

This study analyzed the risk factors for IVIG resistance in Chinese patients with KD and constructed a robust predictive CatBoost model with enhanced interpretation via SHAP analysis. Our predictive model performed well and, for the first time, revealed associations between winter-onset KD, Chlamydia infection, and IVIG resistance. We also externally validated the model, demonstrating its generalizability and robustness. This model may help physicians better manage patients with KD, By enabling the early identification of patients at high risk for IVIG resistance, this model facilitates timely treatment intensification. This could involve the initial co-administration of glucocorticoids with IVIG or the prompt initiation of biologic agents upon early signs of treatment failure, rather than awaiting recrudescent fever to administer a second IVIG course. Such a proactive strategy aims to shorten the total duration of inflammation, which is a key driver of coronary artery lesions, thereby potentially reducing the incidence of this serious complication, avoiding the waste of medical resources, and alleviating the economic burden on families.