We conducted a retrospective study of electronic medical records from patients with KD diagnosed between April 2009 and March 2020 at eight medical institutions affiliated with the Wakayama Prefecture Kawasaki Disease Clinical Research Group (Wakayama Medical University Hospital, Wakayama Rosai Hospital, Naga Public Hospital, Hashimoto Municipal Hospital, Hidaka Hospital, Kinan Hospital, Shingu Municipal Medical Center, and Hannan Municipal Hospital). These data (i.e., the development cohort) were used to create a prediction model for CAA presence after treatment. KD was diagnosed according to the clinical criteria from the fifth and sixth diagnostic guidelines established by the Japan Kawasaki Disease Research Committee (16, 17). The first day of illness was defined as the first day of fever (≥37.5°C).

Patients were excluded from the study if they (1) did not receive IVIG treatment or did not follow our treatment protocol (detailed below) (2), had recurrent KD (3), lacked pre-treatment blood test or follow-up echocardiographic data (4), had incomplete electronic medical records, or (5) did not have recorded height measurements.

All enrolled patients received the following treatment protocol: IVIG (2 g/kg/day for 24 h) + aspirin (30 mg/kg/day) as the first-line therapy; IVIG-resistant patients [persistent or recrudescent fever (axillary temperature ≥37.5°C) 24 h after the first IVIG administration] received a second IVIG dose (2 g/kg/day for 24 h) as a second-line therapy; and patients resistant to the second IVIG administration [persistent or recrudescent fever (axillary temperature ≥37.5°C) after IVIG completion] were treated with oral cyclosporine A (5 mg/kg/day) as a third-line therapy.

This study was approved by the Ethics Committee of the Wakayama Medical University (approval no. 2801) and by the Institutional Review Board or Independent Ethics Committee of each participating medical institution.

Two validation cohorts were enrolled to assess the external and internal validity of the CAA prediction model. The internal validation cohort (VC1) comprised electronic medical records from patients with KD diagnosed at eight participating medical institutions between April 2020 and September 2023. The external validation cohort (VC2) included patients diagnosed at the Department of Pediatrics of Chiba University and Tokyo Women's Medical University Yachiyo Medical Center between January 2017 and December 2019. The same exclusion criteria applied to the development cohort were used for the validation cohorts.

Medical records included patient demographics, clinical characteristics, medications, sequential echocardiographic findings, and clinical laboratory test results prior to the initial treatment. Laboratory tests included white blood cell count, neutrophil percentage, hematocrit, platelet count, aspartate aminotransferase, alanine aminotransferase, total bilirubin, serum albumin, serum sodium, and C-reactive protein (CRP). Variable selection was based on previous KD scoring systems (5–7, 18).

Echocardiographic data were collected for the right coronary artery, left anterior descending artery, and left circumflex artery in two phases: pre-treatment (at KD diagnosis or before IVIG treatment) and convalescent (post-treatment; approximately 4 weeks after KD onset). Left main coronary artery was excluded from the analysis due to frequent anatomical variations (2). Z-scores for these coronary dimensions were retrospectively calculated using the Z-score calculator (version 4.0 Full, LMS_Z_Score) (19). The highest Z-score among the three arteries in each phase was recorded as Zmax. CAA was defined as the post-treatment presence of CAA, based on a Zmax ≥2.5. Consequently, this definition could include patients who already had CAA before treatment (pre-treatment Zmax ≥2.5). The model was developed to predict the post-treatment presence of CAA, encompassing both new-onset and persistent cases.

Continuous variables are reported as medians (interquartile ranges), and categorical variables are presented as frequencies and percentages. Statistical analyses were performed using JMP Pro version 14 (SAS Institute Japan Ltd., Tokyo, Japan). Fisher's exact test was used to analyze categorical variables, whereas the Mann–Whitney U test was employed to analyze continuous variables. For multiple comparisons, a Kruskal–Wallis test was initially performed to detect overall significance between groups. When significant, Dunn's test with Bonferroni correction was applied for pairwise comparisons.

Variables with P < 0.20 in univariate analysis were considered for multivariable modeling. Multivariable logistic regression analysis was applied to assess independent predictors of CAA, with bilateral P values < 0.05 indicating statistical significance. Odds ratios and 95% confidence intervals were calculated. Multicollinearity was assessed based on the variance inflation factor (VIF) of each variable using SPSS version 25 (SPSS Inc., Chicago, IL, USA).

Receiver operating characteristic (ROC) curves were used to determine cutoff values for significant independent predictors of CAA presence after treatment. Predictive variables were assigned scores (1 or 2) based on regression coefficients (RCs), forming a CAA prediction score using pre-treatment clinical data. Sensitivity and specificity were assessed via ROC curves. Model discrimination was evaluated using the area under the ROC curve (AUC), and model performance was additionally assessed using confusion matrices. Furthermore, a sensitivity analysis was conducted using the same method, excluding cases with pre-treatment CAA from the development cohort. The Hosmer–Lemeshow test was used to assess regression model goodness-of-fit using SPSS version 25 (SPSS Inc., Chicago, IL), with P > 0.05 indicating no significant deviation between the model and observed event rates.

In total, 1,230 patients diagnosed with KD at eight medical institutions during the study period were included in the development cohort. The validation cohorts comprised 321 and 413 patients with KD in VC1 and VC2, respectively.

Exclusion criteria led to the removal of patients who did not follow the treatment protocol (n = 63), had recurrent KD (n = 41), lacked echocardiographic follow-up data (n = 32), had incomplete medical records (n = 114), or lacked height measurement records (n = 10). Consequently, 970, 262, and 333 patients in the development cohort, VC1, and VC2 were analyzed, respectively (Figures 1A–C).

In the development cohort, demographic, clinical, treatment, and echocardiographic data at treatment initiation were stratified by CAA status (Table 1). Univariate analysis identified younger age at fever onset, lower hematocrit, platelet count, and albumin levels, and higher CRP and pre-Zmax as significant predictors of CAA (P < 0.20). The VIF values for these variables were all <2.0.

The median time from fever onset to treatment initiation was 5 days, with no significant difference between the CAA and no-CAA groups. In the CAA group, 33 patients (57.8%) showed refractoriness to initial IVIG therapy, and 24 patients (42.1%) had CAA before initial treatment.

Multivariate analysis confirmed that younger age at diagnosis, lower baseline albumin level, and higher pre-Zmax were independent CAA predictors (Table 2A). ROC curve analysis determined the following cutoff values: pre-Zmax ≥1.6, albumin level ≤3.1 g/dL, and age ≤12 months (Supplementary Figure 1). Each cutoff value was significantly associated with CAA: pre-Zmax ≥1.6 (RC: 0.868; P < 0.0001), albumin level ≤3.1 g/dL (RC: 0.699; P < 0.0001), and age ≤12 months (RC: 0.489; P = 0.0008) (Table 2B).

Scores for CAA prediction were assigned based on the ratios of RC values: pre-Zmax ≥1.6 scored 2, whereas albumin level ≤3.1 g/dL and age ≤12 months each scored 1 (Figure 2A). The total score cutoff for predicting CAA, determined using ROC curve analysis, was ≥2, yielding a sensitivity and specificity of 84.2% and 60.8%, respectively (AUC: 0.80) (Figure 2B). The positive predictive value (PPV) was 11.8%, and the negative predictive value was 98.4% (Supplementary Figure 2A). The Hosmer–Lemeshow statistic (P = 0.73) indicated no significant deviation from model fit. Furthermore, when a sensitivity analysis was performed excluding cases with CAA before initial treatment (pre-treatment Zmax ≥2.5), the sensitivity and specificity were 72.9% and 64.8%, respectively (AUC: 0.74) (Supplementary Figure 2B). The RCs for albumin and age increased, whereas that for pre-Zmax decreased (Supplementary Figure 2C). These results suggest that the contribution of pre-treatment coronary Z-score may be relatively llimitited in patients without pre-treatment CAA, although the model appears to retain its overall predictive power.

Clinical characteristics that differed significantly between the development cohort and either validation cohort included the maximum pre-treatment Z-score (pre-Zmax). The incidence of CAA was 5.9% (n = 57) in the development cohort, 4.2% (n = 11) in VC1, and 1.8% (n = 6) in VC2, with VC2 showing a significantly lower incidence compared to the development cohort (Figure 3A).

The model's internal validity was evaluated in VC1, despite significant differences in sex and pre-Zmax between the development cohort and VC1. The model demonstrated excellent discrimination (AUC: 0.88), 72.7% sensitivity, and 89.6% specificity (Figure 3B). The Hosmer–Lemeshow statistic was nonsignificant (P = 0.83), indicating good fit.

External validity was evaluated in VC2, where the model achieved strong discrimination despite the cohort's low CAA incidence. AUC was 0.88, with 83.3% sensitivity and 85.9% specificity (Figure 3C). The Hosmer–Lemeshow statistic (P = 0.25) showed no significant deviation.

Among 1,565 KD cases across three cohorts, 74 remained CAA. Of these, 61 cases (82.4%) were classified as high-risk by our CAA prediction model (Supplementary Figure 2).

In Japan, KD diagnosis and treatment involve not only pediatric cardiologists but also general pediatricians, which could affect the reproducibility of coronary artery Z-scores. There was no standardization of measurement methods between facilities or examiners, and the varying echocardiographic equipment used in this study could also impact consistency. Additionally, in the development cohort, patients were diagnosed according to either the 5th and 6th editions of the Japanese KD guidelines. It is possible that differences in diagnostic criteria may have influenced the date of IVIG administration. However, validation in two independent cohorts indicates that the CAA presence prediction score, including pre-Zmax, remains applicable even when general pediatricians perform coronary artery measurements. Notably, the findings in this study are specific to the treatment protocol used, and coronary Z-scores were calculated using the Coronary Z-Score Calculator (Version 4.0 LMS_Z_Score). All variables in this predictive model were not standardized using Z-score for clinical interpretability. This lack of standardization may limit direct comparisons between RCs. The development cohort for this model included patients with CAA before treatment. The strong correlation between pre-treatment and post-treatment Z-scores may have led to an overestimation of model performance. To evaluate the robustness of the model, we therefore conducted a sensitivity analysis excluding patients with pre-treatment CAA. As expected given the change in cohort structure, the RCs shifted in the sensitivity analysis. However, this does not undermine the overall utility of the model. Nonetheless, this limitation should be acknowledged in interpreting the results. Finally, whether intensive therapy is beneficial for patients predicted to develop CAA remains an open question requiring further investigation in future studies.