This retrospective cohort study enrolled children diagnosed with Kawasaki disease and complicated by coronary artery aneurysm (CAA) at the Guangzhou Women and Children's Medical Center between September 2015 and December 2023. Inclusion criteria were: diagnosis of CAA according to the American Heart Association (AHA) 2017 KD guidelines (coronary artery Z-score ≥2.5). Exclusion criteria were: 1) congenital heart disease or rheumatic disease; 2) missing >20% of key clinical, laboratory, or echocardiographic follow-up data; 3) IVIG or corticosteroid treatment prior to admission; 4) recurrent KD; 5) coronary dilation secondary to cardiomyopathy or congenital coronary anomalies. All children received standard IVIG treatment (a single infusion of 2 g/kg) after onset, along with antiplatelet/anti-inflammatory therapy per guidelines. Following discharge, patients were routinely monitored with echocardiography at approximately 1 week, 1 month, 2 months, and 3 months. For the purpose of this predictive model study, predictor variables were defined using data from the acute phase (pre-IVIG), and the primary outcome was determined solely by the echocardiogram obtained at the 3-month (90 ± 7 days) time point.

Persistent coronary injury was defined according to the AHA 2017 guidelines as the presence of coronary injury on echocardiogram performed ≥90 days after IVIG treatment. Coronary artery dimensions were adjusted for body surface area and expressed as Z-scores, calculated using the regression equation established by Dallaire et al. Coronary injury was graded by Z-score: <2.0 = normal, 2.0–2.4 = dilation, 2.5–4.9 = small aneurysm, 5.0–9.9 = medium aneurysm, ≥10.0 = giant aneurysm. The primary outcome was defined as the persistence of CAA 90 days (±7 days) after disease onset, determined by a Z-score ≥2.5 in any coronary segment at follow-up. For patients with IVIG resistance, the standard rescue therapy at our center was a second dose of IVIG combined with corticosteroids.

The maximum coronary artery Z-score (body surface area-adjusted) was independently assessed by two cardiac sonographers blinded to group assignment. Predictor variables included: Demographics: age, sex, height, weight. Clinical features: KD type (complete/incomplete), IVIG resistance (recurrent fever >36 h after initial IVIG completion), delayed IVIG treatment (>10 days after fever onset), specifically whether the child was an infant [Age_Infant: age <12 months (yes) vs. age ≥12 months (no)]. Coronary parameters: maximum coronary artery Z-score in the acute phase (before IVIG and within 30 days after IVIG). Laboratory indicators: (all collected before IVIG administration): inflammatory markers (CRP, ESR, PCT), complete blood count (WBC, NEU, Hb, PLT), liver and renal function (ALT, AST, Cr), coagulation profile (PT, APTT, Fib), myocardial injury markers (CK-MB). Data were extracted from the structured electronic medical record system and entered independently by two researchers with cross-checking. To ensure methodological consistency across the entire cohort, coronary Z-scores for all patients, including those enrolled from 2015 to 2017, were (re)calculated using the uniform equation specified in the Outcome Definition section (i.e., the Dallaire et al. formula). This process guaranteed that the same diagnostic standard was applied to every participant. All echocardiographic examinations were performed with the child quiet or sedated, in the standard left lateral decubitus position. Examinations followed a systematic assessment protocol, sequentially scanning the parasternal long-axis, short-axis, and apical views. Coronary artery internal diameters were measured at end-diastole at standard proximal locations of the following coronary segments: left main coronary artery (LMCA), left anterior descending artery (LAD), left circumflex artery (LCx), and right coronary artery (RCA). The key anatomical assessment metric for this study was the maximum coronary artery Z-score (ZM), defined as the highest body surface area-adjusted Z-score measured in any of the above coronary segments during the acute phase.

Continuous variables were tested for normality using the Shapiro–Wilk test. Normally distributed data are presented as mean ± standard deviation and compared using the unpaired Student's t-test; non-normally distributed data are presented as median (interquartile range) and compared using the Mann–Whitney U-test. Categorical variables are presented as number (percentage) and compared using the χ² test or Fisher's exact test. Internal validation was performed using 1,000 bootstrap resamples (12). Predictor selection employed LASSO regression (13) combined with bootstrap stability selection. A multivariate logistic regression model was built using the selected predictors, and odds ratios (ORs) with 95% confidence intervals (CIs) were calculated. The model's discrimination, calibration, and clinical utility were evaluated using the area under the receiver operating characteristic (ROC) curve, calibration plots, and decision curve analysis (DCA) It should be noted that the predictor data were largely complete. Specifically, white blood cell count (WBC) was missing in 5 cases (3.7%), and NT-proBNP was missing in 7 cases (5.2%). Data for total bile acids (TBA) and all other variables were fully available, with all missing proportions well below the 20% exclusion criterion. These limited missing data were handled using multiple imputation by chained equations (14). All analyses were performed using R software (version 4.2.2). A two-sided P-value < 0.05 was considered statistically significant.

The retrospective data used in this study were entirely sourced from the Guangzhou Women and Children's Medical Center. The study protocol was reviewed by the Ethics Review Committee of the Guangzhou Women and Children's Medical Center, Guangzhou Medical University. As this is a retrospective study utilizing anonymized data, the committee granted an exemption from ethical review (Approval No: 2025-389B00) and waived the requirement for informed consent.

This study included 135 children with Kawasaki disease (KD) and coronary artery aneurysms (CAA), of whom 80 (59.3%) had persistent CAAs at the 3-month follow-up after intravenous immunoglobulin (IVIG) treatment. Compared to the regressed CAA group, the persistent CAA group demonstrated significantly higher acute-phase maximum coronary Z-scores (ZM) [4.62 (3.72; 5.68) vs. 2.73 (2.61; 3.09)], higher white blood cell counts [19.13 (17.74; 22.76) × 10⁹/L vs. 17.42 (15.33; 19.68) × 10⁹/L], a greater proportion of children aged <12 months (47.50% vs. 18.18%), and a higher rate of delayed IVIG treatment (40% vs. 7.27%). All these between-group differences were statistically significant (P < 0.001). Additionally, significant differences (P < 0.05) were observed in the proportion with C-reactive protein (CRP) >100 mg/L, height, weight, original CRP values, age, and absolute neutrophil count. The detailed baseline characteristics are presented in Table 1.

In clinical prediction model development, the robustness of variable selection is crucial for ensuring model reproducibility. To avoid overfitting or the inclusion of false-positive predictors due to randomness in a single dataset, we employed the rigorous bootstrap-LASSO stability selection method. This approach involved 1,000 bootstrap resamples to simulate data variation. LASSO regression (with 10-fold cross-validation, using λmin as the optimal penalty parameter) was run on each resample. The selection frequency of each variable across all resamples served as a quantitative measure of its predictive stability. The results showed that the maximum coronary Z-score (ZM) was selected in all 1,000 resamples (selection frequency 100%). Age <12 months (Age1) and total bile acid (TBA) demonstrated selection frequencies of 97.9% (979/1,000) and 90.0% (900/1,000), respectively. Variables with selection frequencies ≥80% also included lipase (87.2%, 872/1,000) and delayed IVIG treatment (83.0%, 830/1,000). The specific selection frequencies for all variables are detailed in Table 2.

Using bootstrap-LASSO stability selection (1,000 resamples) and pre-defined variable selection frequency thresholds (100%, 90%, 80%), we constructed three candidate models (Model A, B, C). Internal validation yielded an optimism estimate of 0.031 (95% CI: 0.010–0.061). Model performance was evaluated using optimism-corrected predicted probabilities, while statistical comparisons between models were based on the original predicted probabilities or classification results (detailed in Tables 3, 4).

Based on a comprehensive consideration of performance equivalence, model parsimony, and clinical utility, Model B was selected as the final prediction model. Performance Equivalence: Model B showed no statistically significant difference in discrimination from the more complex Model C (p = 0.389) and demonstrated good calibration. Maximum Parsimony: Model B achieved predictive performance equivalent to the 5-variable model using only 3 variables, potentially enhancing generalizability and clinical ease of use. Clinical Utility: Model B possessed a net clinical benefit nearly identical to Model C (useful range length 0.94 vs. 0.95) and could effectively guide clinical decisions starting from a 5% threshold probability. Based on the regression coefficients of Model B, a nomogram for predicting the risk of persistent CAA was constructed (Figure 4). This tool transforms the complex regression model into an easy-to-use visual scoring system. Clinicians can quickly determine the corresponding individualized risk probability on the chart based on a child's maximum coronary Z-score (ZM), age < 12 months (Age1), and total bile acid (TBA) level. The complete multivariable logistic regression coefficients, standard errors, odds ratios, and 95% confidence intervals for the final prediction model (Model B) are provided in Supplementary Table S1.

This study sets the prediction “starting line” after acute-phase CAA has already occurred. At this specific stage, the focus of clinical decision-making shifts from “whether CAA will occur” to “whether this CAA injury will be persistent”. The maximum coronary Z-score (ZM), as the strongest predictor, directly reflects the severity of acute coronary damage and serves as an important indicator for predicting coronary lesion severity and regression potential, though its predictive power may be influenced by threshold selection, combined indicators, and population differences (15–17). This aligns with the pathophysiological consensus that more significant inflammatory damage to the vascular wall requires longer repair time and is more prone to leaving permanent structural changes (18–20). The inclusion of age <12 months (Age1) in the model suggests that being an infant is independently associated with a higher risk of persistent CAA., This aligns with emerging evidence suggesting that the pathophysiology of KD in infants may lead to more severe vascular injury or impaired repair mechanisms once coronary dilation occurs (21). By quantifying these factors, our study achieves precise risk re-stratification within the population of children with acute-phase CAAs. Furthermore, an important question regarding whether the extent of coronary involvement (i.e., multi-vessel disease) influences prognosis was addressed. We performed a supplementary analysis. Univariate analysis confirmed that the proportion of children with multi-vessel (≥2) involvement was significantly higher in the persistent CAA group than in the regressed group (53.8% vs. 20.0%, p = 0.003). However, when the binary variable “multi-vessel involvement (≥2 vessels)” was included in a multivariable logistic regression model alongside the core predictors (ZM, Age1, TBA), it failed to provide independent predictive information (adjusted odds ratio = 0.57, 95% CI: 0.10–3.37, p = 0.537). This finding carries a dual implication: clinically, it affirms that multi-vessel disease is a clear marker of greater disease severity; for predictive modeling, it indicates that the maximum Z-score (ZM), as a continuous variable, comprehensively captures the key anatomical risk information encompassing both the peak severity and extent of damage. Consequently, for predicting CAA persistence, the “peak intensity” of injury (ZM) constitutes a more robust and stable anatomical predictor than the mere binary classification of multi-vessel involvement. This further supports the selection of ZM as a core predictor.

In children with acute-phase CAAs, elevated total bile acid (TBA) emerged as an independent negative predictor, a finding of significant implications. In this context, TBA may be more than just a marker of liver involvement; its deeper significance potentially reflects the body's response state to severe systemic inflammation and associated metabolic disturbances (22). We hypothesize that during the acute phase of KD, severe systemic inflammation may lead to bile acid metabolism dysfunction and gut-liver axis disruption, resulting in elevated serum TBA levels. The signaling molecule function of elevated TBA might be far more important than its role as a liver injury marker.

Recent research has gradually revealed the crucial role of bile acids as signaling molecules in systemic inflammatory diseases. Their mechanism of action is particularly noteworthy in the context of KD, an acute vasculitis. Bile acids precisely regulate inflammatory signaling through the TGR5-cAMP-PKA axis. TGR5 activation induces PKA-mediated phosphorylation of NLRP3 (at Ser291), promoting its ubiquitination and degradation, thereby inhibiting inflammasome assembly (23). This anti-inflammatory effect is receptor-specific, as it is absent in TGR5-deficient monocytes (24). Substantial evidence confirms the NLRP3/IL-1β pathway as a major driver of KD pathophysiology, with mouse models and patient transcriptome data demonstrating NLRP3 involvement in the development of coronary arteritis (25–28). Similar to autoinflammatory diseases, KD patients exhibit ITPKC gene polymorphisms (affecting endoplasmic reticulum calcium homeostasis) and ER stress, which may activate NLRP3 through yet unidentified mechanisms (28, 29). Single-cell sequencing has revealed altered expression of bile acid metabolism (BAM)-related genes in specific immune cell subsets (e.g., CD8+ T cells) in KD patients, suggesting bile acid signaling may participate in immune regulation (30).

Therefore, measuring TBA in children with acute-phase CAA could help identify “concealed high-risk” individuals who exhibit significant systemic inflammatory responses and poor long-term coronary recovery potential, despite having only mild apparent liver dysfunction, going beyond traditional inflammatory markers like CRP. Consequently, elevated TBA levels in our study might reflect an imbalance between impaired TGR5-mediated anti-inflammatory pathways and excessive NLRP3 inflammasome activation. This imbalance could hinder timely coronary repair, ultimately leading to persistent aneurysms. This also provides new perspectives for future research exploring the modulation of bile acid metabolism or related signaling pathways to improve coronary outcomes in KD.

The primary application of the developed nomogram is risk assessment of children with acute-phase CAA before discharge or shortly after the acute phase. By quickly calculating the risk probability, clinicians can objectively identify high-risk individuals, with particular emphasis oninfants (age <1 year), whom the model identifies as having a significantly elevated inherent risk for persistent injury. To facilitate precise calculation, a detailed score conversion table is provided in Supplementary Table S2, enabling risk estimation without relying solely on the graphical nomogram. For clinical integration, we propose a practical workflow: Risk Quantification: Use the nomogram or Supplementary Table S2 to calculate an individualized risk probability. Risk Stratification: Categorize patients into risk tiers (e.g., low, intermediate, high) based on this probability. The specific probability thresholds defining these tiers may be adapted based on local clinical judgment and resources. Management Triage: Match the risk tier to corresponding management intensity. This early identification allows for the timely enrollment of these children into stricter long-term management pathways, such as more frequent cardiac ultrasound follow-up, prolonged antiplatelet therapy, or even early consideration of secondary prevention medications (e.g., statins) to promote vascular repair (29). Decision curve analysis (DCA) provides objective evidence for this risk-based decision-making. Model B demonstrates net benefit starting from a very low threshold probability of 5%, indicating its clinical utility across a broad spectrum of decision thresholds. This supports the model's role in informing, rather than replacing, clinical judgment. In summary, this nomogram serves as a quantitative decision-support tool. It transforms clinical and laboratory data into an individualized risk estimate, providing an objective basis for tailoring follow-up strategy. Future work may focus on validating specific probability cut-offs for intervention in larger, prospective cohorts.