Early risk stratification for ICU admission in hospitalized children with pertussis: development and internal validation of an admission-time prediction model Authors: Zhi Xia, Ming Chen, Liang Bao, Xiaolan Shu, Xu Zhang, Yifan Luo, Qin Hu, Wenqiang Yang, Manman Xu, Buyun Shi Published: 2026 ABSTRACT BackgroundSevere pertussis in children can deteriorate rapidly, and early identification of patients likely to require intensive care remains challenging. We aimed to develop and internally validate an admission-time prediction model for ICU admission among hospitalized children with PCR-confirmed pertussis.MethodsWe conducted a single-center retrospective prediction-model study based on a hospitalized cohort at a provincial pediatric infectious disease center. Consecutive hospitalized children with PCR-confirmed pertussis between 6 June 2020 and 24 June 2025 were screened. Two children were excluded because admission white blood cell count (WBC), a prespecified core predictor, was unavailable in the retrievable admission-time record, leaving 662 children aged 24 days to 14 years for analysis. The primary outcome was ICU admission, defined a priori as ICU length of stay > 0 days. We prespecified a multivariable logistic regression model using seven admission-time predictors: age (months), sex, symptom duration before admission (days), paroxysmal cyanosis documented at presentation, imaging consolidation/atelectasis, admission WBC scaled per 5 × 109 /L, and any recorded pertussis-containing vaccine dose. Internal validation used 5-fold cross-validation with out-of-fold predictions; model performance was assessed by discrimination (AUC), overall prediction error (Brier score), calibration, and decision curve analysis.ResultsICU admission occurred in 61/662 (9.2%) children. Compared with non-ICU patients, ICU-admitted children were younger (median 2.0 vs 60.0 months) and had higher admission WBC (median 18.7 vs 10.0 × 109 /L). Paroxysmal cyanosis (49.2% vs 6.0%) and imaging consolidation/atelectasis (23.0% vs 10.5%) were more frequent, whereas recorded vaccination was less common (23.0% vs 87.9%). On internal validation, the model showed excellent discrimination (AUC 0.953; 95% CI 0.933–0.972) with low overall prediction error (Brier score 0.045). Calibration was good (intercept −0.017; slope 0.992), and decision curve analysis suggested positive net benefit over clinically relevant threshold probabilities.ConclusionsAn admission-time multivariable model showed good internal performance for predicting ICU admission in pediatric pertussis and may support early risk stratification. External validation is required before routine implementation.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12879-026-13725-5. Background Pertussis remains an important pediatric respiratory infection and continues to cause substantial morbidity and mortality [1, 2]. Recent syntheses of global evidence highlight persistent immune vulnerability in early infancy and the disproportionate burden of severe disease in this age group [3–7]. In China, population-based active surveillance data further indicate that pertussis-related hospitalizations remain a measurable public health problem, underscoring the clinical relevance of hospitalized cohorts [4]. Among hospitalized children, timely identification of those at high risk for deterioration is crucial for decisions regarding monitoring intensity, escalation of respiratory support, pediatric intensive care unit (ICU) resource allocation, and broader hospital planning during periods of increased respiratory infection burden [8]. In critically ill infants, hyperleukocytosis and very young age have been repeatedly linked to poor outcomes, and simplified bedside scores incorporating these factors have been proposed [9, 10]. In addition, prior ward-versus-ICU studies and more recent pediatric cohorts have identified younger age, cyanosis or respiratory distress, lack of diphtheria-tetanus-acellular pertussis (DTaP) vaccination, and leukocytosis as important severity signals [11–14]. However, several gaps limit translation into routine practice. First, “severe pertussis” has been defined heterogeneously across studies, which complicates cross-center comparability and external validation. Second, some models incorporate variables not available at admission or focus on PICU-only cohorts, limiting their utility for admission-time triage in general pediatric wards. Third, many reports emphasize discrimination (AUC), but provide limited information on calibration and clinical utility, which are essential for bedside decision-making [15–18]. To address these issues, we developed and internally validated an admission-time prediction model for ICU admission among hospitalized children with PCR-confirmed pertussis. We selected ICU admission as an objective operational endpoint (ICU length of stay > 0 days) to enhance reproducibility and facilitate comparisons across settings where thresholds for specific interventions may vary. Using routinely available predictors available at admission, we evaluated discrimination, calibration, overall prediction error, and clinical utility to support early risk stratification and pragmatic bedside implementation. Methods Study design and setting We conducted a single-center retrospective prediction-model study based on a hospitalized cohort at Maternal and Child Health Hospital of Hubei Province, a designated pediatric infectious disease care center (Children’s Infectious Disease Prevention and Control Center of Hubei Province). The hospital inpatient pediatric system includes the pediatric ICU and general pediatric wards. Consecutive hospitalized children with PCR-confirmed pertussis admitted between 6 June 2020 and 24 June 2025 were screened according to admission dates recorded in the electronic medical record (EMR). Data source, variable extraction, and missing-data handling Clinical and laboratory data were extracted from the hospital EMR system for analysis. The extracted dataset included demographic information, admission-time clinical features, imaging-related variables, laboratory results on admission, vaccination history recorded in the medical chart, ICU length of stay, and diagnosis fields used for comorbidity sensitivity analyses. Age and sex were obtained from structured demographic fields in the EMR. Symptom duration before admission and paroxysmal cyanosis at presentation were abstracted from the chief complaint and admission record. Vaccination status was abstracted from the recorded vaccination history in the medical chart. Imaging consolidation/atelectasis was determined from the first available chest radiograph or chest CT obtained at admission, or from the most recent outpatient imaging immediately before hospitalization if it was used for admission assessment. Admission WBC was defined as the first available in-hospital WBC count measured at admission. Among the 664 screened children, 2 were excluded because admission WBC, a prespecified core predictor, was unavailable in the retrievable admission-time record. Although partial blood-count information from outside hospitals was available before admission in these children, these incomplete external results were not substituted for the prespecified admission WBC variable. No implicit imputation or default coding assumptions were used for the prespecified primary-model predictors. Predictors were extracted only when explicitly documented in the relevant admission-time record, structured assessment form, or first available admission investigation. The observed age range in the screened cohort was 24 days to 14 years. Case definition, eligibility, and outcome Pertussis was confirmed in all included patients by a positive oropharyngeal swab polymerase chain reaction (PCR) test for Bordetella pertussis together with compatible respiratory symptoms at presentation. Eligible participants were hospitalized children in the inpatient pediatric system who met this laboratory-confirmed case definition. No additional exclusions were made on the basis of underlying comorbidities. The analytic cohort therefore consisted of patients with complete prespecified admission-time predictors used in the primary model. Participant flow is shown in Fig. 1.Fig. 1Study flow diagram. Among 664 hospitalized children with PCR-confirmed pertussis, 2 were excluded because admission WBC was unavailable in the retrievable admission-time record, leaving 662 in the analytic cohort. ICU admission was defined as ICU length of stay > 0 days Study flow diagram. Among 664 hospitalized children with PCR-confirmed pertussis, 2 were excluded because admission WBC was unavailable in the retrievable admission-time record, leaving 662 in the analytic cohort. ICU admission was defined as ICU length of stay > 0 days The primary outcome was ICU admission, defined a priori as ICU length of stay > 0 days. This operational definition was chosen to provide a reproducible and clinically actionable endpoint for early risk stratification. Candidate predictors To support true early risk stratification, we prespecified predictors that are available at admission. The primary model included age (months, continuous), sex (male vs female), symptom duration before admission (days, continuous), paroxysmal cyanosis documented at presentation or on admission assessment (yes/no), imaging consolidation/atelectasis (yes/no), admission WBC (×109 /L, continuous), and vaccination status (any recorded pertussis-containing vaccine dose: yes/no). For interpretability, admission WBC was scaled per 5 × 109 /L (WBC5 = WBC/5). Binary variables were coded as 1 for yes/present and 0 for no/absent. Statistical analysis Baseline characteristics were summarized by ICU admission status. Continuous variables are presented as median (IQR, Q1–Q3) and were compared using the Mann–Whitney U test. Categorical variables are presented as counts (percentages) and were compared using the χ2 test or Fisher’s exact test, as appropriate. We developed a multivariable logistic regression model to estimate the probability of ICU admission using the prespecified admission-time predictors. Continuous predictors (age, symptom duration before admission, and WBC) were modeled as linear terms in the primary model. Results are reported as adjusted odds ratios (aORs) with 95% confidence intervals (CIs). Multicollinearity was assessed using variance inflation factors (VIFs), with all predictors showing low collinearity (VIF < 2). The full model equation and coefficients for individual risk calculation are provided in Supplementary Methods S1 and Supplementary Table S1. Internal validation was performed using 5-fold cross-validation. For each fold, the model was trained on 80% of the data and evaluated on the held-out 20%, yielding out-of-fold (OOF) predicted probabilities for all patients. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Overall prediction error was assessed using the Brier score. Calibration was evaluated using a calibration plot and quantified by calibration intercept and slope. Clinical utility was assessed using decision curve analysis (DCA) based on OOF predictions. Threshold-based classification performance at prespecified probability thresholds of 10 and 20% is reported in Supplementary Table S2. Additional sensitivity analyses were performed to address concerns regarding comorbidity, interaction, non-linearity, vaccination-age eligibility, and model robustness. Major underlying comorbidity was defined as clearly documented congenital heart disease, major airway structural abnormality, asthma/chronic airway disease, neurologic or aspiration-prone underlying disorder, or immunologic disorder. Because this variable was ascertained from diagnosis fields in routinely archived discharge documentation rather than uniformly available admission-time records, it was not treated as a primary admission-time predictor and was examined only in sensitivity analyses (Supplementary Tables S3 and S4). We also explored an age-by-cyanosis interaction; assessed potential non-linearity for age and admission WBC using cubic spline terms with likelihood-ratio testing (Supplementary Table S5; Supplementary Fig. S1); repeated the vaccination analysis among children who had reached vaccine-eligible age (Supplementary Table S5); and fitted ridge-penalized and class-weighted logistic regression models as robustness checks for modest events-per-variable (EPV) and class imbalance (Supplementary Table S5). All analyses were performed in Python (version 3.11.2) using pandas, NumPy, SciPy, statsmodels, and scikit-learn; figures were generated using Matplotlib. Reporting guideline This study was reported with reference to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guideline for prediction model development and internal validation and the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline for observational studies. Results Study population A total of 664 hospitalized children with PCR-confirmed pertussis were screened. Two children were excluded because admission WBC was unavailable in the retrievable admission-time record, leaving 662 children in the analytic cohort with complete prespecified admission-time predictors. ICU admission (ICU length of stay > 0 days) occurred in 61/662 (9.2%) children, whereas 601/662 (90.8%) were not admitted to ICU (Fig. 1). A total of 62/662 children (9.4%) had a clearly documented major underlying comorbidity based on diagnosis-field review (Supplementary Table S3). In sensitivity analyses that additionally accounted for, or excluded, major underlying comorbidity, the directions of the primary predictors and overall model performance were materially unchanged (Supplementary Table S4). Baseline characteristics by ICU admission Baseline admission-time characteristics are summarized in Table 1. Children admitted to the ICU were markedly younger than those not admitted (median age, 2.0 months [IQR, 2.0–4.0] vs 60.0 months [IQR, 8.0–84.0]; p < 0.001) and had substantially higher admission WBC (18.7 × 109 /L [IQR, 12.0–35.8] vs 10.0 × 109 /L [IQR, 7.4–13.6]; p < 0.001). Paroxysmal cyanosis at presentation was far more frequent in the ICU group (30/61, 49.2% vs 36/601, 6.0%; p < 0.001), and imaging consolidation/atelectasis was also more common (14/61, 23.0% vs 63/601, 10.5%; p = 0.007). Recorded pertussis-containing vaccination was less common among ICU-admitted children (14/61, 23.0% vs 528/601, 87.9%; p < 0.001). Major underlying comorbidity was also more common in the ICU group (23/61, 37.7% vs 39/601, 6.5%; p < 0.001). Symptom duration before admission did not differ between groups (median 10.0 vs 10.0 days; p = 0.788), and male sex showed a non-significant trend toward higher ICU admission (63.9% vs 51.6%; p = 0.088).Table 1Baseline characteristics by ICU admission status (analytic cohort)CharacteristicNon-ICU (n = 601)ICU (n = 61)p valueAge, months60.0 (8.0–84.0)2.0 (2.0–4.0)<0.001Symptom duration before admission, days10.0 (6.0–15.0)10.0 (6.0–15.0)0.788WBC on admission (×109 /L)10.0 (7.4–13.6)18.7 (12.0–35.8)<0.001Male sex310 (51.6)39 (63.9)0.088Paroxysmal cyanosis at presentation36 (6.0)30 (49.2)<0.001Imaging consolidation/atelectasis63 (10.5)14 (23.0)0.007Vaccinated (any dose recorded)528 (87.9)14 (23.0)<0.001Major underlying comorbidity39 (6.5)23 (37.7)<0.001Values are median (IQR, Q1–Q3) or n (%). Continuous variables were compared using the Mann–Whitney U test; categorical variables were compared using the χ2 test or Fisher’s exact test, as appropriateMajor underlying comorbidity was defined for sensitivity analysis only and was not included in the primary nomogram Table 1Baseline characteristics by ICU admission status (analytic cohort)CharacteristicNon-ICU (n = 601)ICU (n = 61)p valueAge, months60.0 (8.0–84.0)2.0 (2.0–4.0)<0.001Symptom duration before admission, days10.0 (6.0–15.0)10.0 (6.0–15.0)0.788WBC on admission (×109 /L)10.0 (7.4–13.6)18.7 (12.0–35.8)<0.001Male sex310 (51.6)39 (63.9)0.088Paroxysmal cyanosis at presentation36 (6.0)30 (49.2)<0.001Imaging consolidation/atelectasis63 (10.5)14 (23.0)0.007Vaccinated (any dose recorded)528 (87.9)14 (23.0)<0.001Major underlying comorbidity39 (6.5)23 (37.7)<0.001Values are median (IQR, Q1–Q3) or n (%). Continuous variables were compared using the Mann–Whitney U test; categorical variables were compared using the χ2 test or Fisher’s exact test, as appropriateMajor underlying comorbidity was defined for sensitivity analysis only and was not included in the primary nomogram Baseline characteristics by ICU admission status (analytic cohort) Values are median (IQR, Q1–Q3) or n (%). Continuous variables were compared using the Mann–Whitney U test; categorical variables were compared using the χ2 test or Fisher’s exact test, as appropriate Major underlying comorbidity was defined for sensitivity analysis only and was not included in the primary nomogram Multivariable predictors of ICU admission In the prespecified multivariable logistic regression model, younger age, paroxysmal cyanosis at presentation, imaging consolidation/atelectasis, and higher admission WBC (per 5 × 109 /L increase) remained independently associated with ICU admission, whereas recorded vaccination was associated with lower odds of ICU admission (Table 2; Supplementary Fig. S2). Sex and symptom duration before admission were not independently associated with ICU admission.Table 2Multivariable logistic regression for ICU admission (adjusted odds ratios)PredictorAdjusted OR95% CIp valueAge (per 1 month increase)0.920.86–0.990.025Male sex1.050.48–2.290.906Symptom duration before admission (per 1 day)0.960.92–1.010.109Paroxysmal cyanosis at presentation5.172.33–11.49<0.001Imaging consolidation/atelectasis4.821.65–14.040.004Admission WBC (per 5 × 109 /L increase)1.811.45–2.25<0.001Vaccinated (any dose recorded)0.280.12–0.650.003Adjusted ORs were derived from the prespecified 7-predictor admission-time multivariable logistic regression model Table 2Multivariable logistic regression for ICU admission (adjusted odds ratios)PredictorAdjusted OR95% CIp valueAge (per 1 month increase)0.920.86–0.990.025Male sex1.050.48–2.290.906Symptom duration before admission (per 1 day)0.960.92–1.010.109Paroxysmal cyanosis at presentation5.172.33–11.49<0.001Imaging consolidation/atelectasis4.821.65–14.040.004Admission WBC (per 5 × 109 /L increase)1.811.45–2.25<0.001Vaccinated (any dose recorded)0.280.12–0.650.003Adjusted ORs were derived from the prespecified 7-predictor admission-time multivariable logistic regression model Multivariable logistic regression for ICU admission (adjusted odds ratios) Adjusted ORs were derived from the prespecified 7-predictor admission-time multivariable logistic regression model Internal validation and model performance On internal validation using 5-fold cross-validation with OOF predictions, the model demonstrated excellent discrimination (AUC 0.953; 95% CI 0.933–0.972) with low overall prediction error (Brier score 0.045) (Table 3; Fig. 2). Calibration was good, supported by a near-zero calibration intercept (−0.017) and a slope close to 1.0 (0.992), consistent with the calibration plot (Table 3; Fig. 3). Decision curve analysis showed positive net benefit across clinically relevant threshold probabilities (Fig. 4).Table 3Model performance (internal validation with 5-fold cross-validation)MetricEstimate95% CIAUC (5-fold CV)0.9530.933–0.972Brier score (5-fold CV)0.0450.035–0.057Calibration slope (OOF)0.9920.773–1.212Calibration intercept (OOF)−0.017−0.450–0.416AUC and Brier score 95% CIs were derived from bootstrap resampling of OOF predictions; calibration confidence intervals were obtained from the calibration modelFig. 2Receiver operating characteristic (ROC) curve based on out-of-fold predictions from 5-fold cross-validationFig. 3Calibration plot based on out-of-fold predictions from 5-fold cross-validation (isotonic calibration curve and deciles with 95% CI)Fig. 4Decision curve analysis (DCA) based on out-of-fold predictions from 5-fold cross-validation. Net benefit is shown across threshold probabilities for the prediction model compared with treat-all and treat-none strategies. Vertical reference lines indicate prespecified risk thresholds of 10 and 20% used for threshold-based performance reporting Table 3Model performance (internal validation with 5-fold cross-validation)MetricEstimate95% CIAUC (5-fold CV)0.9530.933–0.972Brier score (5-fold CV)0.0450.035–0.057Calibration slope (OOF)0.9920.773–1.212Calibration intercept (OOF)−0.017−0.450–0.416AUC and Brier score 95% CIs were derived from bootstrap resampling of OOF predictions; calibration confidence intervals were obtained from the calibration model Model performance (internal validation with 5-fold cross-validation) AUC and Brier score 95% CIs were derived from bootstrap resampling of OOF predictions; calibration confidence intervals were obtained from the calibration model Receiver operating characteristic (ROC) curve based on out-of-fold predictions from 5-fold cross-validation Calibration plot based on out-of-fold predictions from 5-fold cross-validation (isotonic calibration curve and deciles with 95% CI) Decision curve analysis (DCA) based on out-of-fold predictions from 5-fold cross-validation. Net benefit is shown across threshold probabilities for the prediction model compared with treat-all and treat-none strategies. Vertical reference lines indicate prespecified risk thresholds of 10 and 20% used for threshold-based performance reporting Additional model checks and sensitivity analyses Additional model checks showed no evidence supporting retention of an age-by-cyanosis interaction term in the primary model, and age showed no meaningful non-linearity. Admission WBC showed some evidence of non-linearity, but the overall predictive performance changed only modestly. In age-eligible vaccination analysis, the vaccination coefficient was attenuated and no longer statistically significant. Ridge-penalized regression yielded performance nearly identical to the primary model, whereas class-weighted regression did not improve overall performance (Supplementary Table S5; Supplementary Fig. S1). Clinical utility and model presentation Decision curve analysis indicated that using the model would yield greater net benefit than treat-all or treat-none strategies across clinically relevant threshold probabilities (Fig. 4). Threshold-based classification performance at prespecified risk thresholds of 10% and 20% is reported in Supplementary Table S2. A nomogram was constructed to facilitate bedside estimation of individual risk using the final primary model (Fig. 5).Fig. 5Nomogram for predicting ICU admission in hospitalized children with PCR-confirmed pertussis. Points are summed to total points and mapped to the predicted probability of ICU admission. The nomogram was derived from the final multivariable logistic regression model including age (months), sex, symptom duration before admission (days), paroxysmal cyanosis documented at presentation, imaging consolidation/atelectasis, admission WBC scaled per 5 × 109 /L, and vaccination status Nomogram for predicting ICU admission in hospitalized children with PCR-confirmed pertussis. Points are summed to total points and mapped to the predicted probability of ICU admission. The nomogram was derived from the final multivariable logistic regression model including age (months), sex, symptom duration before admission (days), paroxysmal cyanosis documented at presentation, imaging consolidation/atelectasis, admission WBC scaled per 5 × 109 /L, and vaccination status Discussion In this cohort of 662 hospitalized children with PCR-confirmed pertussis, we developed an admission-time prediction model for ICU admission using seven routinely available variables. Younger age, paroxysmal cyanosis at presentation, imaging consolidation/atelectasis, and higher admission WBC emerged as the strongest risk signals, whereas recorded vaccination was associated with lower odds of ICU admission. Internal validation showed excellent discrimination, low overall prediction error, good calibration, and positive net benefit across clinically relevant thresholds. These findings are clinically coherent and align with the established profile of severe pediatric pertussis. Across recent cohorts and reviews, the burden of critical illness remains concentrated in early infancy, and the children who deteriorate most often show a combination of age-related vulnerability, respiratory compromise, and marked leukocytosis [9, 12–14, 19, 20]. In that context, the strong association between cyanosis and ICU admission in our cohort is biologically plausible. Cyanosis is an early bedside marker of impaired gas exchange or ventilatory compromise in infants with pertussis, and it has repeatedly appeared alongside other severity signals in hospitalized or critically ill cohorts [11–13, 20]. The retained WBC effect is likewise consistent with prior reports linking leukocytosis and hyperleukocytosis to severe disease, pulmonary vascular complications, and poorer outcomes [10–14, 21, 22]. Where our study adds practical value is in how these signals were assembled into a truly admission-time tool. Several prior studies of severe or critical pertussis have relied on peak in-hospital laboratory values, treatment-course variables, or ICU-only populations [21–23]. Those approaches are useful for understanding disease severity, but they are less suitable for early triage in general pediatric wards or emergency settings. We deliberately restricted the primary model to variables that are typically available at the first hospital assessment. That design choice keeps the model close to real clinical workflow and makes the resulting nomogram easier to implement as a bedside probability aid rather than a retrospective severity label. We also chose imaging consolidation/atelectasis rather than auscultatory findings as the representative marker of pulmonary involvement. This was a pragmatic decision. In routine care, auscultatory signs such as crackles or rales are partly subjective and may vary across examiners, whereas chest imaging findings are recorded in radiology reports and are generally more reproducible in retrospective extraction [24]. For the same reason, we did not build the model around later complications such as pulmonary hypertension or exchange-transfusion decisions, although these features are clinically important in malignant pertussis [9, 21]. Likewise, pathogen co-infection may influence clinical phenotype, but such information is not consistently available at the first assessment and was therefore beyond the scope of the primary admission-time model [25]. Our goal was not to maximize model complexity, but to preserve early usability. The vaccination coefficient deserves cautious interpretation. In the primary model, recorded vaccination was associated with lower odds of ICU admission. However, many ICU-admitted children in this cohort were very young infants, and some would not yet have reached the age of the first scheduled pertussis-containing vaccine dose. Consistent with this concern, the vaccination coefficient was attenuated and no longer statistically significant in the age-eligible sensitivity analysis. This pattern suggests that the vaccination variable in the primary model should be interpreted as a risk-stratification signal within this hospitalized cohort rather than as a simple causal estimate of vaccine effect. This interpretation is also consistent with recent pediatric pertussis cohorts in which severe or ICU-level disease clustered in very young infants, many of whom were below or near the age of the first scheduled dose [12–14]. The additional model checks further support the stability of the primary model. We did not find evidence supporting retention of an age-by-cyanosis interaction term, and age showed no meaningful non-linearity. Admission WBC did show some evidence of non-linearity, but the gain from a more flexible spline specification was small in practical terms. Ridge-penalized regression yielded performance nearly identical to the primary model, which is reassuring given the modest events-per-variable ratio, while class-weighted regression did not provide a clear advantage. Taken together, these analyses suggest that the primary seven-predictor model is reasonably robust, even though it is intentionally parsimonious. Major underlying comorbidity was addressed in sensitivity analyses rather than incorporated into the primary admission-time model because these diagnoses were not uniformly ascertainable at the first clinical contact. In routine practice, some important underlying conditions, such as congenital heart disease or airway structural abnormalities, may only be confirmed after additional diagnostic work-up during hospitalization and are therefore more reliably captured in archived discharge documentation than in admission-time records. Accordingly, we considered major underlying comorbidity inappropriate as a primary bedside predictor, but suitable for sensitivity analyses. Nevertheless, the main findings were materially unchanged in those analyses, which supports the robustness of the primary model. This study has several limitations. First, it was a single-center retrospective study, and ICU admission practices may vary across hospitals. Second, the outcome was ICU admission rather than a strictly biologic severity endpoint; although operationally reproducible, it may partly reflect local admission thresholds. Third, the number of ICU events was limited relative to the number of predictors, and the ICU/non-ICU groups were unbalanced, although ridge-based and class-weighted sensitivity analyses did not materially alter the overall conclusions. Fourth, diagnosis-field ascertainment of underlying comorbidity is imperfect, and some residual clinical heterogeneity cannot be excluded. Fifth, recorded vaccination status in young infants is partly coupled to age eligibility. Finally, external validation and prospective impact evaluation are still required before routine implementation. In conclusion, we developed and internally validated an admission-time multivariable logistic regression model to estimate ICU admission risk among hospitalized children with PCR-confirmed pertussis. The model demonstrated strong internal performance and may support early risk stratification in routine care. External validation and prospective impact evaluation are warranted before routine implementation. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary material 1 Supplementary material 2 Supplementary material 3 Supplementary material 4 Supplementary material 1 Supplementary material 2 Supplementary material 3 Supplementary material 4