Kawasaki disease (KD) is an acute systemic vasculitis mainly affecting children under 5 years of age and is characterized by fever, polymorphic rash, conjunctival congestion, changes in oropharyngeal mucosa, changes in the peripheral extremities, and cervical lymphadenopathy [1]. Although the etiology of KD remains unknown, the most common view is that an unknown stimulus triggers the activation of an immune-mediated inflammatory cascade in genetically susceptible children [2]. Coronary artery aneurysms (CAAs), the most serious complication of KD, occur in approximately 25 % of untreated children, making this disease the leading cause of acquired heart disease among children in developed countries [3], [4]. The standard therapy for KD patients is high-dose intravenous immunoglobulin (IVIG) and aspirin, which reduces the risk of CAA to 3 %–6% [5]. However, up to 20 % of KD patients with recurrent or persistent fever at least 36 h after initial treatment are resistant to IVIG [6]. Existing research suggests that the emergence of IVIG resistance may be closely related to a variety of mechanisms. These include the influence of genetic factors, an inadequate response of the immune system to immunoglobulins, and abnormal responses of the immune system such as cytokine storms, imbalances between T helper cell 17 and regulatory T cell, the diminished function of immunosuppressive receptors, and abnormalities in immunoregulatory pathways [5]. Consequently, these IVIG-resistant patients face a higher risk of developing CAA and are more likely to have an echocardiographic manifestation of CAA in the acute phase of the disease [7]. In patients with acute KD who are at high risk for IVIG resistance or who have a high risk of CAA, IVIG in combination with adjuvant glucocorticoids or IVIG in combination with other nonglucocorticoid immunomodulatory immunosuppressive agents is conditionally recommended as initial therapy, rather than IVIG therapy alone [8].

Thus, identifying predictors of IVIG-resistant KD is crucial for early diagnosis and enabling personalized treatment strategies. Researchers around the world are continuously working to develop various predictive scoring systems or predictive models. Investigators have reported several scoring systems predicting IVIG resistance, such as the Kobayashi, Egami, and Sano scoring systems in Japan [9], [10], [11] and the San Diego scoring system in the United States [12]. However, scoring systems were insufficiently robust to be clinically useful in ethnically diverse populations [13]. Chinese researchers have also developed some predictive models for IVIG resistance. For example, the Formosa score, used to detect IVIG resistance in Taiwan, analyzed data from 181 KD patients, achieving a sensitivity of 90.9 % and a specificity of 81.3 % [14]. Huang et al. proposed a nomogram with good discriminatory power (AUC of 0.75) to predict IVIG resistance in hospitalized children in Suzhou, China [15]. Yang et al. constructed a scoring criterion: total bilirubin > 20 μmol/L (5 points), C-reactive protein (CRP) ≥90 mg/L (3 points), serum sodium <135 mmol/L (3 points), percentage of neutrophils ≥70 % (2.5 points), albumin <35 g/L (2.5 points); and when the score ≥6 was classified as a high-risk group, the model had a sensitivity of only 56 % and specificity of 79 % [16]. This prediction tool was useful for the early screening of high-risk IVIG resistance in KD in Beijing. Additionally, Tan et al. devised a predictive model for IVIG-resistant KD in Chongqing [17]. It can be seen that different regions have different prediction models and there is no uniformly applicable scoring model [18]. Hence, we aimed to develop a new IVIG resistance prediction model for KD patients in Shanghai to help clinicians identify high-risk cases early and administer prompt interventions.

In recent years, machine learning (ML) algorithms have played a broad role in medicine. By leveraging big data sets and predictive models, ML enables clinicians to diagnose, predict, and treat with greater confidence, leading to personalized medicine [19]. In addition, ML plays a role in drug development, improving the efficiency of drug discovery and pharmacokinetic prediction, and helping to predict patients' response to drugs. In the fields of hematology, oncology, and histopathology, ML is used to predict patient survival outcomes using clinical images and data [20]. The applications of ML in medicine go far beyond these. To summarize, ML algorithms show great potential in the field of medical prediction applications. This study employed four ML algorithms: random forest (RF), support vector machine (SVM), stepwise regression analysis, and least absolute shrinkage and selection operator (LASSO). Each of these algorithms has unique advantages and application scenarios. Firstly, RF is an ensemble learning algorithm based on decision trees, which handles non-linear relationships, mitigates overfitting by constructing multiple trees, and provides variable importance for dependent variable prediction, thereby exhibiting superior predictive performance and aiding clinical decision-making [21], [22], [23]. Secondly, SVM, a supervised learning algorithm, can find an appropriate hyperplane in a high-dimensional feature space, effectively separating different classes of data points, and demonstrating strong classification capability even with limited training samples [24]. Thirdly, stepwise regression analysis is a method for linear feature selection, which simplifies the model and improves the predictive accuracy of the dependent variable by iteratively adding or removing variables [25]. Fourthly, LASSO is a linear model that constrains the model's complexity by introducing an L1 regularization term, leading some coefficients to become zero and removing predictor variables that do not contribute to the model, which performs excellently in handling high-dimensional datasets and situations with multicollinearity [26], [27].

We aimed to select the most convenient, concise, and efficient predictive model for IVIG resistance among these ML algorithms. The model would be visualized through a dynamic nomogram, aiding physicians in predicting the risk of IVIG resistance quickly and intuitively, enabling timely interventions to reduce the likelihood of CAA and improve the long-term prognosis.