This study leveraged clinical data sourced from three tertiary medical institutions in China: Wuxi People's Hospital, Wuxi Second People's Hospital, and Tengzhou Central People's Hospital. Inclusion criteria encompassed: (1) infants aged 0–12 months at enrollment; (2) undergoing vascular lesion screening at birth or during infancy prompted by physical examination findings or clinical symptoms; (3) availability of comprehensive perinatal data, including gestational age, birth weight, and Apgar scores; (4) detailed maternal-infant records, comprising pregnancy-related complications, conception mode, and history of drug exposure; (5) documented family consent for longitudinal follow-up, alongside either in-hospital birth registration or complete follow-up documentation. Exclusion criteria comprised: (1) confirmed diagnoses of syndromic vascular anomalies, including but not limited to PHACE syndrome, Sturge-Weber syndrome, or CLOVES syndrome, as well as concurrent non-hemangioma vascular malformations; (2) known chromosomal aberrations or major structural defects such as trisomy 21 or severe cardiac and cerebral malformations; (3) extreme prematurity (gestational age <28 weeks) or extremely low birth weight (<1,000 g); (4) presence of profound immunodeficiency or antecedent neoplastic conditions, including congenital immunodeficiency syndromes; (5) mortality or attrition within 12 months postpartum. In this study, the case group comprised infants and young children with clinically, radiologically, and pathologically confirmed IH. The control group was drawn from contemporaneous infants undergoing routine check-ups or clinical visits at the same hospitals, all of whom were systematically screened to exclude IH, other vascular tumors, and congenital vascular malformations. Furthermore, individuals with evident infections, immunological disorders, metabolic diseases, or other severe systemic conditions were excluded to ensure that the control group accurately represented a population of “healthy infants without IH.” This retrospective investigation received ethical approval from the institutional review boards of all participating centers, with informed consent requirements duly waived.

A comprehensive set of 40 clinical variables encompassing diverse domains was systematically collected to enable an exhaustive risk assessment for infantile hemangiomas. These variables comprised: Demographic and parental factors including infant sex, small-for-gestational-age (SGA) status, parental age, and mode of delivery; Perinatal and obstetric history such as maternal American Society of Anesthesiologists (ASA) score, maternal smoking and alcohol consumption, history of miscarriage, placental abnormalities, paternal smoking and alcohol use, family history of hemangiomas, multiple gestation, prematurity, and low birth weight. Maternal comorbidities and antenatal conditions included hormone therapy during pregnancy, intrauterine infection, gestational hypertension, gestational diabetes, maternal anemia, and umbilical cord complications. Neonatal conditions and congenital anomalies encompassed Apgar scores, congenital heart disease (CHD), and gestational age classification. Laboratory biomarkers incorporated serum albumin (ALB), C-reactive protein (CRP), serum amyloid A (SAA), vascular endothelial growth factor (VEGF), interleukin-6 (IL-6), tumor necrosis factor-alpha (TNF-α), and neutrophil-to-lymphocyte ratio (NLR). Tumor characteristics and clinical manifestations were also recorded, including hemangioma subtype, age at onset, lesion size and morphology, anatomical location, presence of complications, and lesion count. The principal outcome measure was the occurrence of infantile hemangioma. In this study, biomarkers including VEGF, CRP, and SAA were derived from blood tests conducted during routine postnatal check-ups and early clinical visits, collected within 1–6 months after birth. The majority of samples were obtained prior to the clinical confirmation of IH or during the early stage of the lesion, typically at the time when a suspicious lesion was first identified. Recognizing that SAA and CRP are acute-phase reactants susceptible to elevation during acute infections, which could confound study outcomes, we implemented rigorous measures during study design and data collection to mitigate this effect. All enrolled infants underwent comprehensive clinical evaluation at the time of sampling, with those exhibiting overt signs of infection (e.g., fever, respiratory or urinary tract infections) systematically excluded. Beyond SAA and CRP, additional infection-related laboratory indices, including white blood cell count and neutrophil proportion, were incorporated to identify and omit cases with potential active infection. Sample collection was further confined to infants without a recent history of acute infection (e.g., within the preceding two weeks) to minimize interference. Collectively, these measures ensured that observed variations in SAA and CRP more faithfully reflected the immune-inflammatory milieu pertinent to infantile hemangioma.

Variables exhibiting a missing rate below 5% were designated as having low missingness, whereas those with missing rates ranging from 5% to 30% were considered to possess moderate to high missingness. Two complementary strategies were employed to address these gaps. For variables with low missingness, simple imputation was implemented: continuous variables were imputed using the median, and categorical variables were imputed using the mode (most frequent category). This method, confined to scenarios of minimal missingness, aimed to preserve sample integrity and was subsequently benchmarked against multiple imputation outcomes in sensitivity analyses. For variables with moderate to high missingness, multiple imputation was undertaken. Binary variables were imputed via logistic regression models, wherein the probability distribution of missing values was inferred from available predictors, followed by stochastic sampling to retain intrinsic inter-variable correlations. Multicategorical variables were addressed using multinomial logistic regression, concurrently estimating the probability of each mutually exclusive category and imputing missing entries through probabilistic sampling. This approach preserved the original data distribution, mitigated bias, enhanced the plausibility of imputations, and consequently bolstered the predictive robustness of downstream models.

The diagnosis of infantile hemangioma was ascertained through a comprehensive clinical framework, predominantly grounded in meticulous physical examination and augmented by high-resolution imaging modalities as warranted (18–20). Initial assessments were performed by seasoned pediatricians or dermatologists, drawing upon key clinical hallmarks including lesion onset, rapid proliferative behavior, characteristic coloration ranging from vivid red to bluish-purple, soft and elevated consistency, blanching response, and the quintessential triphasic pattern of proliferation, plateau, and involution.

For lesions exhibiting classical phenotypes, diagnosis was rendered on clinical grounds alone. In instances of atypical morphology, suspected deep tissue infiltration, or to discriminate from alternative vascular anomalies such as vascular malformations or angiosarcoma, color Doppler ultrasonography was employed to appraise lesion depth, margins, internal structure, and hemodynamic features. Lesions located in anatomically intricate regions—such as the orbit, cervical area, oropharynx, or visceral organs—or those demonstrating extensive involvement, were further evaluated using magnetic resonance imaging (MRI) to precisely delineate lesion boundaries and their spatial relationships with adjacent tissues. Cases presenting equivocal imaging findings underwent multidisciplinary review by senior radiologists. Definitive diagnoses were established through consensus by no fewer than two senior pediatric specialists, synthesizing clinical presentation, imaging data, and longitudinal follow-up, thereby ensuring strict adherence to diagnostic criteria for infantile hemangioma among all enrolled subjects.

This study utilized SPSS and R software to develop and systematically evaluate a clinical prediction model through the following steps: 1.

Data preprocessing.

A total of 1,466 infants and young children were enrolled in the study, including 81 cases of hemangioma (Table 1, Supplementary Table S1 and Figure 1). The cohort was predominantly female (71.6%), with males comprising 28.4%. The median age at onset was 18.0 months. Within the group, 33.3% were SGA, 27.2% were firstborns, and 35.8% had a neonatal Apgar score below 7. Multiple gestations accounted for 53.1% of cases, 44.4% were born prematurely, and 60.5% had low birth weight. Nuchal cord occurrence was observed in 16.0% of neonates. Consistent with prior research, hemangiomas were classified morphologically into focal, segmental, indeterminate, and multifocal subtypes. Focal hemangiomas predominated, representing 61.7% of cases, followed by indeterminate (24.7%) and segmental (12.3%) types; multifocal lesions were rare, present in only 1.2% of patients. The majority of children (77.8%) presented with solitary lesions, whereas 22.2% exhibited multiple lesions. Lesion distribution was highest on the head and neck (62.96%), followed by the face (16.0%), trunk (11.1%), extremities (7.4%), and perineum (2.5%). Regarding complications, 65.4% of patients were complication-free. Among those affected, ulceration was the most frequent (22.2%), followed by auditory or airway obstruction (4.9%), vision impairment (2.5%), secondary infection (2.5%), and bleeding (2.5%). Notably, these complications frequently co-occurred with ulceration. The internal dataset included 818 participants, of whom 48 had infantile hemangioma (IH). The external dataset comprised 648 participants, including 33 IH cases. A comparison of the features is presented in Table 2. The internal dataset was randomly divided into training and testing sets at a 7:3 ratio, with their characteristics compared in Table 3. The original dataset utilized in this study is provided in Supplementary Table S2. To ensure the reproducibility and transparency of this study, all source code used—including scripts for data preprocessing, model construction, performance evaluation, and SHAP analysis—is available at the permanent access link (https://www.jianguoyun.com/p/DWh9chMQl-GKDBjEj-sFIAA).

Both univariate and multivariate logistic regression analyses identified several independent risk factors for infantile hemangioma development, including gestational diabetes mellitus, mode of delivery, multiple pregnancy, preterm birth, low birth weight, Apgar score, and elevated levels of VEGF, CRP, and SAA (P < 0.05) (Table 4). These results highlight the multifactorial etiology of infantile hemangioma, implicating perinatal factors alongside inflammatory and angiogenic biomarkers in its pathogenesis.

To further refine the risk factor profile, we employed four classical machine learning algorithms—XGBoost, RF, SVM, and KNN—for feature selection. The overlap of top-ranked features across all models consistently identified multiple pregnancy, preterm birth, low birth weight, and elevated VEGF, CRP, and SAA levels as the strongest predictors of infantile hemangioma (Figures 2A–D). This machine learning–augmented strategy corroborated the logistic regression findings and enhanced the robustness of the key predictive variables. The hyperparameters of the four machine learning models were optimized via grid search, with XGBoost set as colsample_bytree = 1, learning_rate = 0.3, max_depth = 4, min_child_weight = 4, n_estimators = 20, reg_lambda = 0.5, and subsample = 1; RF as criterion = gini, max_depth = None, max_features = sqrt, min_impurity_decrease = 0.0, min_samples_leaf = 1, min_samples_split = 2, and n_estimators = 100; SVM as C = 1.0, gamma = scale, kernel = rbf, max_iter = 50, probability = True, and tol = 0.001; and KNN as algorithm = auto, leaf_size = 10, n_neighbors = 4, p = 2, and weights = uniform.

ROC curve analysis demonstrated that the XGBoost model exhibited superior predictive performance in both the training and validation cohorts, achieving an AUC of 0.952 in the training set and 0.935 in the validation set—the highest among the four evaluated machine learning algorithms (Table 5 and Figures 3A–C). These elevated AUC values underscore the model's excellent discriminative ability to differentiate between high- and low-risk infants, reflecting a high degree of predictive accuracy. Calibration curves for XGBoost, RF, SVM, and KNN revealed strong agreement between predicted probabilities and observed outcomes, indicating good calibration and reliable probability estimation across all models. Additionally, DCA assessed the clinical utility of each model, demonstrating that across a spectrum of threshold probabilities, all models provided greater net clinical benefit compared to “treat-all” or “treat-none” approaches (Figure 3D). Notably, XGBoost delivered the most favorable clinical decision support, highlighting its promise for personalized risk stratification in infantile hemangioma. To rigorously evaluate model generalizability, 10-fold cross-validation was performed within the internal cohort. Specifically, 245 cases (30.0%) were randomly assigned as a test set, while the remainder were used for training and cross-validation. This approach minimized sampling bias and enhanced robustness by averaging performance across multiple data partitions. In cross-validation, XGBoost achieved the highest overall performance with a validation AUC of 0.9438 ± 0.0484, test set AUC of 0.8366, and accuracy of 0.8943 (Figures 4A–C). By comparison, the RF model showed a validation AUC of 0.8510 ± 0.1334, test AUC of 0.8353, and accuracy of 0.8415; SVM yielded a validation AUC of 0.8326 ± 0.1362, test AUC of 0.6827, but the highest accuracy at 0.9472; and KNN demonstrated a validation AUC of 0.8466 ± 0.1243, test AUC of 0.8064, and accuracy of 0.8780. These results collectively emphasize the consistent superiority of XGBoost in terms of AUC, accuracy, and stability, establishing it as the most effective algorithm for predicting high-risk infantile hemangioma. External validation using an independent cohort further corroborated the model's generalizability, with XGBoost achieving an AUC of 0.870 (Figure 4D), confirming robust predictive capability on unseen data. The Kolmogorov–Smirnov (KS) curve demonstrates a clear separation between the cumulative distribution curves of IH cases and non-IH controls, with a pronounced maximum vertical distance (KS value), indicating the model's efficacy in distinguishing high-risk from low-risk samples. The confusion matrices for both the training and testing sets reveal that true positives (TP) and true negatives (TN) markedly exceed false positives (FP) and false negatives (FN), underscoring the model's robust classification performance and accuracy. Parallel coordinates plots exhibit consistent line patterns across samples for different features, effectively illustrating each feature's contribution to model predictions and highlighting distinctions in the multi-feature space between high-risk and low-risk samples, with prediction patterns remaining stable and devoid of notable anomalies (Figures 5A–D).

The SHAP summary plot (Figure 6) offers a lucid visualization of the principal risk factors associated with infantile hemangioma, ranking them according to their relative contribution to the model's output. The analysis identified SAA level, low birth weight, VEGF level, multiple pregnancy, preterm birth, and CRP level as the most influential predictors.

To further assess the model's clinical interpretability and applicability, we examined individual prediction outcomes for four representative patients using SHAP force plots (Figures 7A–D). These plots illuminate patient-specific high-risk contributors and quantify their respective impact magnitudes. Patient 1: The model predicted a high probability (0.82) of developing infantile hemangioma, primarily driven by elevated CRP and SAA levels, preterm birth, and multiple pregnancy, indicating a high-risk profile. Patient 2: Predicted risk was low (0.06), with minor contributions from CRP levels and low birth weight, suggesting a limited cumulative effect of risk factors in this case. Patient 3: The model estimated a probability of 0.04, where CRP levels and multiple pregnancy were the main contributors, indicating a low overall risk. Patient 4: Predicted probability was 0.05, with CRP level and multiple pregnancy as the key contributing factors. Although classified as low risk, ongoing monitoring of these variables may be advisable. These individualized explanations demonstrate the capacity of the XGBoost model combined with SHAP analysis to enable precision risk stratification, thereby supporting nuanced and informed clinical decision-making.