A retrospective cohort study was conducted, including 51 cases of MIS-C with post–SARS-CoV-2 infection. All registered cases involved children aged 7 months to 17 years who were hospitalized for inpatient treatment in healthcare facilities with varying levels of resource availability. The verification of MIS-C diagnosis was performed according to the CSTE/CDC diagnostic criteria (2022) (12) with exclusion of active acute infection in patients.

The analytical procedures were performed using the statistical software EZR v.1.54. Descriptive statistics for potential prognostic factors were presented for quantitative variables were presented as median and interquartile range (Me (IQR) or as mean ± standard deviation (M ± SD), depending on normality. Normality of data distribution was assessed using the Shapiro–Wilk W-test. Qualitative (binary) variables were expressed as absolute and relative frequencies (n, %). Intergroup comparisons were conducted using the Student's t-test or Mann–Whitney U-test for continuous variables, and the χ²-test or Fisher's exact test for categorical variables. A two-tailed p-value < 0.05 was considered statistically significant.

The following analysis involved univariate logistic regression (ULR), with a p < 0.1 defined as the threshold for statistical significance.

Variable selection for the multivariable model was performed using a forward stepwise approach based on the Akaike information criterion (AIC).

We assessed multicollinearity using the variance inflation factor (VIF) and considered values below 5 as acceptable.

To develop severity scoring models for MIS-C, multivariate logistic regression (MLR) was applied, and the independent association of each predictor with the outcome was deemed statistically significant when p < 0.05. The predictive ability of the model was evaluated using receiver operating characteristic (ROC) curve analysis. The area under the curve (AUC) was computed to determine the model's discriminative accuracy. Odds ratios (ORs) with corresponding 95% confidence intervals (CIs) were subsequently calculated to estimate the strength of association and to visualize the probability of classification errors. The optimal cut-off point for continuous variables was determined using the Youden index. Specifically, ROC analysis for each continuous biomarker was performed with PICU admission (PICU+ vs. PICU–) as the reference outcome, and the threshold maximizing the Youden index (sensitivity plus specificity minus 1) was selected. The resulting cut-offs were used to dichotomize variables for subsequent regression analyses and score construction (e.g., ESR ≥ 22 mm/h; IL-6 ≥ 310 pg/mL).

Internal validation of the prognostic models was performed using bootstrap resampling with 1,000 iterations. Model performance was evaluated across replications based on the area under the ROC curve (AUC) and the Brier index, and their mean values with 95% confidence intervals were estimated.

For practical implementation of the developed models, two scoring systems were designed: the MIS-C – Severity Score Basic (MIS-C-SS-Basic) for healthcare facilities with limited resources, and the MIS-C – Severity Score Advanced (MIS-C-SS-Advanced) for institutions with enhanced access in laboratory tests, especially to interleukin-6 (IL-6) measurement. Each factor incorporated into the corresponding model was assigned an index value determined by its weighting coefficient (β), indicating its relative impact in each model.

The study cohort included 51 children; 24 (47%) required PICU admission. The median PICU length of stay lasted 5.5 (3.75–6.25) days (p < 0.001; W = 0.565). Patients admitted to the PICU had a longer total hospital stay, with a statistically significant difference — 15 (11–17) vs. 11 (9–14) days (p = 0.011).

The median age of children who required PICU admission was 6.00 (2.00–10.00) years, compared with 6.55 (4.00–10.00) years in those treated without intensive care. In the PICU group, boys predominated, with a male-to-female ratio of 2.4:1 (17 (70%) vs. 7 (29%)). In contrast, the sex ratio was nearly balanced among children who did not require intensive care — 0.9:1 (13 (48%) vs. 14 (52%)).

The sex distribution did not differ significantly between groups (p = 0.374 and p = 0.154, respectively). A comprehensive baseline description of demographic, clinical, and major laboratory parameters stratified by PICU admission is provided in the Supplementary Table S1 regardless of statistical significance. In the following sections, we focus on clinically relevant differences and key predictors associated with PICU requirement.

Univariable comparisons between PICU+ and PICU− groups are summarized in Tables 1, 2. Table 1 presents baseline clinical characteristics, including ultrasound findings, whereas Table 2 presents baseline laboratory parameters. These analyses provide the descriptive context for subsequent univariable logistic regression assessing associations with PICU admission. Variables with p < 0.10 in preliminary comparisons were entered into ULR as a screening step to avoid excluding potentially clinically relevant predictors.

As shown in Table 1, several baseline clinical characteristics were compared between PICU+ and PICU− groups, allowing identification of clinically relevant differences to be explored in univariate analyses.

To identify prognostic factors, we used a ULR analysis. For patients with MIS-C who required PICU admission (n = 24) the dependent variable was defined as Y = 1, and Y = 0 for those managed outside the PICU (n = 27). Results of the ULR are shown in Table 3.

As shown in Table 3, ULR models identified several clinical and laboratory parameters that were significantly associated with the likelihood of a severe course of MIS-C. A higher risk of severe MIS-C was observed in patients presenting with involvement of six organ systems (p = 0.035, OR = 4.80, 95% CI [1.12–20.61]), hepatomegaly (p = 0.002, OR = 8.05, 95% CI [2.12–30.60]), and elevated IL-6 levels exceeding 310 pg/mL (p < 0.001, OR = 16, 95% CI [3.68–69.60]). Clinically significant predictors also included pharyngitis (p = 0.013, OR = 15.60, 95% CI [1.80–135.00]) and the presence of free fluid in the abdominal cavity (p = 0.040, OR = 5.80, 95% CI [1.10–11.70]). Hypoalbuminemia (< 32.0 g/L) was associated with an increased risk of severe disease (p = 0.033, OR = 0.09, 95% CI [0.01–0.83]), indicating its potential diagnostic value.

Thus, the results indicate that multisystem involvement, combined with pronounced inflammatory activity and hepatomegaly, represents an important predictor of adverse disease course in MIS-C and may be integrated into risk stratification models.

The MLR model was developed based on variables that demonstrated discriminatory ability in predicting the risk of intensive care unit admission and the overall severity of MIS-C.

Considering the need for tiered stratification according to the level of healthcare resources, two operational models were developed. The Advanced model included involvement of four organ systems, hepatomegaly, and interleukin-6 (IL-6) levels above 310.0 pg/mL, designed for centers with access to routine and rapid IL-6 testing. In contrast, the Basic model is proposed for institutions without such access and comprises the presence of free abdominal fluid (excluding ascites), hepatomegaly, and a dichotomous ESR variable (threshold ≥ 22 mm/h). Both models demonstrated a high degree of internal stability, with all included predictors retaining statistical significance (p < 0.05). A comparative visualization of the two resulting models is presented in Figure 1.

Figure 1 presents two ROC curves that markedly deviate from the line of random classification (AUC = 0.5), indicating good discriminative ability of both models. It should be noted that the stability of the prognostic model for MIS-C severity was strongly influenced by the laboratory marker included in it. Incorporation of interleukin-6 increased the AUC value to 0.936 (95% CI [0.86–1.00]), whereas substitution with ESR reduced it to 0.826 (95% CI [0.71–0.94]). The summarized odds ratios and 95% confidence intervals for both models are presented in Table 4.

Table 4 presents the results of the multivariate logistic regression analysis for two predictive models of MIS-C severity — MIS-C-SS-Advanced and MIS-C-SS-Basic.

Both models demonstrated significant associations between clinical and laboratory predictors. In the MIS-C-SS-Advanced model, the most powerful predictor was interleukin-6 (IL-6) level (p < 0.001, OR = 63.5, 95% CI [5.5–733.0]), followed by involvement of four organ systems (p = 0.033, OR = 19.1, 95% CI [1.3–288.0]) and hepatomegaly detected by ultrasound (p = 0.006, OR = 14.5, 95% CI [2.1–98.6]). In the MIS-C-SS-Basic model, which did not include IL-6, all three predictors remained statistically significant — hepatomegaly (p = 0.005, OR = 8.3, 95% CI [1.9–36.4]), presence of free abdominal fluid (p = 0.014, OR = 9.1, 95% CI [1.6–52.7]), and elevated ESR (p = 0.043, OR = 6.6, 95% CI [1.1–41.2]).

The MIS-C-SS-Advanced model demonstrated the highest discriminatory ability (AUC = 0.936). An optimal cut-off value of 0.68 provided high sensitivity (92%), allowing the detection of nearly all severe cases, with a specificity of 89%. Given these parameters, the model is unlikely to miss patients who truly require intensive care, although a small proportion of children with moderately severe disease may be classified as false positives. However, this characteristic is particularly valuable in clinical settings, where the primary priority is to avoid missing critically ill patients.

The MIS-C-SS-Basic model, which incorporated ESR instead of interleukin-6, showed a slightly lower yet still clinically meaningful predictive accuracy (AUC = 0.826, cut-off = 0.31, sensitivity = 87.5%, specificity = 70.4%).

The Basic model showed lower discriminative capacity than the Advanced version; nevertheless, its use remains justified in healthcare facilities with limited resources, where IL-6 testing is either unavailable or time-consuming and therefore not clinically relevant in critical conditions requiring urgent decision-making and patient triage.

Thus, both models can be regarded as reliable tools for predicting the risk of severe MIS-C. However, the Advanced model should be prioritized in centers with access to comprehensive laboratory testing, while the Basic model represents a practical adaptation for resource-limited settings, ensuring a balance between diagnostic accuracy and accessibility.

Considering the relative sample size (n = 51), internal validation of the predictive models was performed using the bootstrap resampling method (bootstrap validation, BV) with B = 1,000 iterations. To assess model performance, the bias-corrected area under the ROC curve (AUC) was estimated, calibration plots were constructed, and the Brier score was calculated. Visualization of the bootstrap-calibrated plots is presented in Figure 2.

The results of the bootstrap validation are summarized in Table 5.

Table 5 summarizes the main statistical results of both models after bootstrap validation with 1,000 iterations. The comparative analysis demonstrates good overall predictive performance of both models, though with differing levels of accuracy.

The Advanced model exhibited enhanced performance in terms of its discrimination capacity (AUC = 0.94 [0.86–0.99]) and overall accuracy (lower Brier score = 0.086). However, the Hosmer–Lemeshow test on the development dataset revealed evidence of imperfect group calibration (p = 0.001), which should be carefully considered during external validation. In comparison, the Basic model retained an acceptable level of predictive validity (higher Brier score = 0.165, but HL p = 0.143), indicating adequate calibration across risk deciles, which is particularly relevant in settings with restricted access to laboratory diagnostics. Hence, both models may be implemented in routine clinical practice, although the Advanced model is preferable when interleukin-6 measurement is available.

To facilitate clinical implementation by physicians, both Advanced and Basic logistic regression models were transformed into point-based scoring systems. This approach serves as an optimization strategy to enhance simplicity and speed of use in routine and emergency clinical settings.

Each predictor variable with independent prognostic value was assigned a specific score using the relative β-coefficient method. The smallest absolute β-coefficient within each final model was taken as the reference unit (Basic: ESR, β = 1.89; Advanced: hepatomegaly, β = 2.67), and all other coefficients were scaled and rounded proportionally to the nearest integer. The reference predictor (|β|) within each model was set to 1 point. Points for the remaining predictors were obtained by dividing each β-coefficient by this reference value and rounding the result to the nearest integer. Detailed scoring system and risk strata are presented in Figure 3.

As shown in Figure 3, the MIS-C-SS-Basic model included three predictors: ESR ≥ 22 mm/h, hepatomegaly, and the presence of free fluid in the abdominal cavity, each assigned 1 point. The MIS-C-SS-Advanced model comprised IL-6 ≥ 310 pg/mL (2 points), involvement of four organ systems, and hepatomegaly (1 point each). IL-6 was assigned 2 points because it had the largest regression coefficient in the Advanced model (β = 4.15) within the relative β-coefficient approach.

To define low-, intermediate-, and high-risk categories, each possible total score was mapped to the corresponding predicted probability of PICU admission derived from the underlying logistic regression model (using the model intercept and coefficients). Based on the monotonic increase in predicted risk across point totals and to provide three clinically pragmatic triage strata, cut-offs were defined as follows: for MIS-C-SS-Basic (0–3 points), low risk = total 0–1 (= 17.0–36.1%), intermediate risk = total 2 (= 60.9%), and high risk = total 3 (= 81.1%); for MIS-C-SS-Advanced (0–4 points), low risk = total 0–1 (= 8.6–26.4%), intermediate risk = total 2 (= 57.6%), and high risk = total 3–4 (= 83.8–95.1%). These thresholds are shown in Figure 3.

The optimal cut-off value for the Basic model was 2 points, providing a sensitivity of 95.8% and specificity of 81.5%. For the model including IL-6, the optimal threshold was 3 points, yielding a sensitivity of 87.5% and specificity of 92.6%.

Additionally, both scoring systems were used to calculate the percentage risk of severe MIS-C and the need for intensive care unit admission. The probabilities were derived using the logistic equation, constructed from the intercept (β0) and the smallest significant β-coefficient as the step per point. The data presented in Figure 3 demonstrate a direct relationship between the total score and the predicted risk of severe MIS-C. According to the results, each patient has a non-zero baseline probability of severe disease within the applied model, even in the absence of any risk factor – 17.0% for MIS-C-SS-Basic and 8.6% for MIS-C-SS-Advance.

As the total score increases, the predicted risk rises accordingly. At 2 points, the probability exceeds the 50% threshold for both models (60.9% for Basic and 57.6% for Advance). At 3 points, the risk of severe disease reaches 81.1% in the Basic model and 83.8% in the Advanced model. The maximum score of 4 points corresponds to the highest predicted risk of 95.1%.

Considering the characteristics of the laboratory parameters included in the developed scoring systems – interleukin-6 as a more specific but less accessible biomarker, and ESR as a less specific marker with slower dynamic changes—it appears reasonable to apply a dynamic use of the scoring approach for appropriate patient management. Figure 4 illustrates the proposed algorithm for implementing the scoring systems and the corresponding clinical decision-making pathway.

The Figure 4 illustrates a stepwise approach for implementing the MIS-C-SS-Advanced and MIS-C-SS-Basic scoring systems in clinical practice. The algorithm comprises four sequential stages: initial patient assessment, risk stratification into three categories, with preliminary branching depending on availability of laboratory diagnostics, followed by subsequent clinical management options.

Understanding the percentage risk of requiring intensive care enables the use of these scoring systems as a tool for guiding clinical and strategic decision-making in patient triage. The Basic model can be applied in any healthcare setting; however, its lower discriminative ability should be considered, implying a higher likelihood of false-positive or false-negative assessments. In contrast, incorporating interleukin-6 levels into the evaluation enhances the accuracy and reliability of clinical decision-making.

The proposed scoring systems allow for a rapid assessment of the risk of PICU admission in patients with MIS-C. Their high sensitivity with a minimal number of available variables makes them suitable for initial patient triage in resource-limited clinical settings. In cases where IL-6 measurement is available, the Advanced version of the score provides greater discriminative performance.