Preoperative hemorrhagic risk stratification in pediatric moyamoya disease: a multi-institutional propensity score-matched analysis

Abstract

Background:

Pediatric hemorrhagic moyamoya disease (MMD) is rare, and currently, no risk model exists for predicting preoperative bleeding. We aimed to develop a nomogram to predict the preoperative bleeding risk in children with MMD.

Methods:

We retrospectively analyzed data from 1350 children diagnosed with MMD from January 2004 to December 2022 at our institution. After applying propensity score matching (PSM), 392 patients were selected for analysis, comprising 98 with hemorrhagic MMD and 294 with non-hemorrhagic MMD. The cohort was divided into training and internal validation cohorts. To construct the nomogram, variable selection was performed using the least absolute shrinkage and selection operator (LASSO), and the model was externally validated with an independent cohort of 70 children. We utilized multivariate logistic regression to determine odds ratios and 95% confidence intervals for preoperative bleeding risk. A predictive nomogram was then developed from the logistic model, with polynomial equations to quantify risk. The model’s effectiveness was evaluated using receiver operating characteristic curves, calibration plots, and decision curve analyses (DCAs). Inflection points for continuous variables were identified using restricted cubic spline (RCS) analysis.

Results:

The LASSO model demonstrated superior discriminative performance compared to six alternative models, achieving area under the curve values of 91.5% in the training cohort, 78.4% in the internal validation cohort, and 91.2% in the external validation cohort. Based on variables selected through the LASSO model, we developed a nomogram incorporating three critical factors: age at onset (P = 0.001), anterior choroidal artery grades 1 (P = 0.047) and 2 (P < 0.001), and posterior communicating artery grades 1 (P = 0.002) and 2 (P = 0.032). Calibration plots indicated strong concordance between predicted and observed outcomes across both training and validation cohorts (Hosmer–Lemeshow P = 0.503), affirming the model’s accuracy. Additionally, DCA highlighted the nomogram’s clinical utility by effectively identifying patients at high risk. RCS analysis revealed age 8 as a pivotal inflection point (P < 0.05), marking a significant increase in the risk of preoperative bleeding beyond this age.

Conclusion:

The nomogram demonstrated high accuracy in predicting preoperative bleeding risk in pediatric patients with MMD. This predictive accuracy may enhance preoperative evaluation by surgeons, allowing for more proactive intervention and intensified monitoring of children at elevated risk of bleeding, thereby improving patient outcomes.

Key messages

What is known about this topic

Preoperative bleeding is a critical determinant influencing treatment strategies and long-term outcomes in pediatric patients with hemorrhagic moyamoya disease (MMD). Current prognostic models are predominantly tailored to adult MMD populations, leaving a significant gap in predictive tools for preoperative bleeding risk in children. To date, no validated risk model exists specifically for pediatric patients, underscoring the need for targeted research in this area.

What the study adds

This study addresses this gap by developing and validating a nomogram to predict the risk of preoperative bleeding in pediatric MMD. The model incorporates three key predictors: age at onset, anterior choroidal artery (grades 1 and 2), and posterior communicating artery (grades 1 and 2). Notably, the risk of preoperative bleeding increases with age at onset, with a distinct inflection point observed at 8 years. This nomogram provides a robust, evidence-based tool for risk assessment in pediatric MMD, offering insights into the interplay of clinical and vascular factors.

How this study might affect research, practice, or policy

The introduction of this validated nomogram has the potential to significantly impact clinical practice by enabling more precise risk stratification and individualized preoperative management for pediatric MMD patients. By identifying high-risk patients, clinicians can optimize treatment strategies to mitigate bleeding risks and improve outcomes. Furthermore, this study lays the groundwork for future research into the mechanisms underlying preoperative bleeding in pediatric MMD and informs policy development aimed at enhancing patient care and reducing adverse events.

Introduction

Moyamoya disease (MMD) is a rare cerebrovascular condition characterized by progressive stenosis and occlusion of the internal carotid arteries and the development of abnormal collateral vessels at the base of the brain, leading to diminished cerebral blood flow and an increased risk of ischemic and hemorrhagic strokes^[1]. Although its exact etiology remains unclear, genetic and environmental factors are believed to contribute to MMD development^[2]. MMD can affect individuals of all age groups, with an estimated first peak incidence between ages 5 and 10 years and a second peak in the fourth decade of life^[3].

The clinical presentation of pediatric MMD varies widely, with recurrent transient ischemic attacks (TIA), infarctions, and no symptoms being the most common. However, hemorrhagic presentations, although less common in children compared to adults^[4], with an incidence rate of 0%–9.3%^[5,6], can result in catastrophic outcomes, making a thorough understanding of risk factors crucial for optimal management. One area of particular concern is the risk of preoperative bleeding, which can complicate treatment plans and affect surgical outcomes. Therefore, accurate prediction and management of preoperative bleeding in pediatric MMD are critical^[7]. Although several clinical and radiographic factors for hemorrhagic MMD have been identified, including sex and dilatation of the anterior choroidal artery (AChA) and posterior communicating artery (PComA)^[8], the precise predictive value of these factors remains unclear. Surgical revascularization, including direct and indirect procedures, remains the mainstay for managing MMD. These interventions can improve blood flow in the affected areas of the brain and reduce the risk of recurrent bleeding in children with hemorrhagic MMD^[5]. However, the preoperative period poses unique risks, including the potential for bleeding due to fragile collateral vessels and altered hemodynamics. Identifying factors that contribute to an increased risk of preoperative bleeding could better inform clinical practices and perioperative management strategies in these vulnerable patients.

Despite the clinical significance, few studies have focused on predicting preoperative bleeding in pediatric MMD, and the existing prognostic models are largely based on adult MMD populations. Furthermore, existing studies often suffer from small sample sizes and lack comprehensive multi-institutional data, which are essential for drawing robust conclusions. Additionally, traditional statistical techniques often fail to capture the complex, nonlinear relationships inherent in clinical data.

HIGHLIGHTS

• This study developed a nomogram to predict the risk of the preoperative bleeding in children with moyamoya disease.

• The age at onset at the inflection point with an increased preoperative bleeding risk was established as 8 years.

• The predictive accuracy may enhance preoperative evaluation by surgeons monitoring pediatric patients with a high risk of preoperative bleeding.

In response to these limitations, our study aims to harness the power of machine learning to develop a predictive nomogram tailored for estimating the risk of preoperative bleeding in pediatric MMD. By employing seven machine-learning frameworks, this study integrates advanced computational techniques with clinical insights to provide a nuanced understanding of bleeding risks. These models will be developed and validated using a multi-institutional propensity score-matched dataset, ensuring robustness and real-world applicability of the findings. The ultimate goal of this research is to produce a reliable, user-friendly nomogram that can be seamlessly integrated into clinical workflows, offering healthcare providers a valuable tool for preoperative risk stratification and individualized patient management, and guiding clinicians in optimizing preoperative assessment and improving surgical outcomes for this delicate patient population.

Materials and methods

Ethical review

This study complied with the Declaration of Helsinki and was approved by the local Institutional Review Board and Ethics Committee. An exemption from obtaining informed consent was granted because of the retrospective nature of the data. All patient data were anonymized prior to analysis, with identifiers removed to ensure confidentiality. Data governance adhered to China’s Personal Information Protection Law and institutional protocols, including secure encrypted storage and restricted access limited to authorized study personnel.

Patient selection and study design

The training cohort for this study was collected from our hospital between January 2004 and December 2022, while the external validation cohort was sourced from a different medical center over the same period. The diagnosis was based on the criteria of the Research Committee on MMD (spontaneous occlusion of the circle of Willis) of the Ministry of Health, Labour and Welfare, Japan, and the 2021 Guideline Committee of the Japan Stroke Society^[9]. The inclusion criteria were as follows: (i) diagnosis of MMD based on digital subtraction angiography (DSA), (ii) age <18 years, and (iii) availability of complete data on the predictive factors studied. The exclusion criteria were as follows: (i) age ≥18 years, (ii) incomplete or missing data on the predictive factors studied, (iii) presence of underlying vascular or bleeding disorders, and (iv) presence of other medical conditions, such as central nervous system tumors, severe brain trauma history, previous craniotomy, and genetic diseases. In addition, those with Moyamoya syndrome were excluded from this study. The original data were generated and stored at our hospital. Data supporting these findings are available from the corresponding author upon request. This study was reported in line with the STROCSS criteria^[10].

Patients were categorized into three groups: the training cohort, internal validation cohort, and external validation cohort. The training cohort, comprising 70% of the pediatric patients with MMD (N = 274) from a retrospective study at our hospital, aimed to develop a scoring system to predict the risk for preoperative bleeding in pediatric patients with MMD. The internal validation cohort, consisting of the remaining 30% of the pediatric patients with MMD (N = 118) from our hospital, validated the diagnostic effectiveness of this scoring system. The external validation cohort involved 70 prospectively enrolled pediatric patients with MMD at another medical center, independent of those at our hospital, for further assessment of the predictive model.

Propensity score matching

To address the differing demographic characteristics between pediatric MMD patients with preoperative bleeding and those without, this study employed optimal fixed-ratio propensity score matching (PSM) to identify comparable patient groups. PSM analysis was conducted to mitigate potential bias in research outcomes by minimizing confounding effects associated with mismatched baseline characteristics^[11]. Propensity scores were calculated using a logistic regression model that incorporated demographic and clinical covariates, including sex, initial symptoms (e.g. TIA, infarction, hemorrhage subtypes), age at onset, hemorrhage type (intraparenchymal, intraventricular, or subarachnoid hemorrhage), laterality (unilateral or bilateral involvement), preoperative modified Rankin Scale (mRS) score, family history of cerebrovascular disease, and Suzuki angiographic stage (I–VI). The binary response variable for PSM was defined as the presence or absence of AChA dilation, a critical radiographic marker previously associated with hemorrhagic risk in MMD.

The matching procedure utilized a 1:3 greedy nearest-neighbor algorithm with a caliper width of 0.2 standard deviations of the logit propensity score, implemented via the R package MatchIt (version 4.5.0). This caliper width was selected based on established recommendations to balance matching precision and retained sample size while minimizing bias. To ensure covariate balance between the AChA dilation and non-dilation groups, post-matching equilibrium was assessed using standardized mean differences (SMDs) for continuous variables and absolute percentage differences for categorical variables, with an SMD threshold of <0.1 indicating adequate balance. Sensitivity analyses, including quantile-quantile (Q-Q) plots and Kolmogorov–Smirnov tests, were performed to validate the distributional equivalence of propensity scores across matched cohorts. This rigorous approach achieved balanced distributions of all predefined covariates, thereby reducing selection bias and enabling robust comparisons of preoperative bleeding outcomes.

Cohort and variable definition

Pediatric patients with MMD were stratified into training and internal validation cohorts in a 7:3 ratio through the “createDataPartition” function in R, guaranteeing a random distribution of outcome events across both cohorts. The training cohort was used for variable selection and model construction, whereas the validation cohort was employed to evaluate the findings derived from the training cohort. The following demographic and clinical data were obtained: sex, initial symptoms, age at onset, laterality, preoperative mRS, family history, Suzuki stage, PComA grade, and AChA grade. The algorithm was configured with a predefined random seed (seed = 1234) to ensure reproducibility, and the partitioning process iteratively verified event rate consistency between cohorts using chi-square tests (P > 0.05).

The definitions of the items evaluated in this study are established in the field. The onset of symptoms was determined based on the clinical presentation of the initial attack or MRI results indicating infarction. Pediatric patients exhibiting signs of epilepsy and TIA were considered to have infarction if a subsequent MRI study indicated infarction. If the TIA evolved into epilepsy, it was classified as epilepsy. To evaluate the impact of preoperative neurological function on cerebral hemorrhage, mRS, with age-specific modifications^[12], was used to classify patients based on the severity of their neurological deficits. The scale values ranged from 0 to 6, with higher values reflecting intensifying levels of disability. An mRS score of ≥2 reflects poor neurological function, indicating at least moderate disability. Specifically, patients with a score of 2 on the mRS can typically walk independently but require help with everyday tasks. The baseline mRS score was determined based on the patient’s initial hospitalization.

Radiological evaluation

Infarction diagnosis relied on MRI findings of the cerebral hemisphere, where it was identified as a recent lesion showing heightened intensity on T2- and diffusion-weighted images, accompanied by reduced intensity on the apparent diffusion coefficient image. The diagnosis of intracranial hemorrhage was based on high-density intracranial lesions discovered on initial computed tomography. Although other symptoms appear before bleeding, this definition still applies. Two experienced neurosurgeons assessed the preoperative Suzuki stage^[13] and used the staging of the hemorrhagic hemisphere if there were disparities between the two cerebral hemispheres. AChA and PComA were graded based on DSA according to previously published methods^[14,15]. Two senior radiological specialists performed the AChA and PComA grading scale. Based on the AChA grading scale, the absence of dilation is classified as grade 0. A dilated AChA within the choroidal fissure corresponds to grade 1, whereas a dilated AChA that extends beyond the choroidal fissure is classified as grade 2. Grade 3 is designated for situations in which the AChA is no longer visible because of occlusion of the internal carotid artery. In the context of PComA, a non-dilated state was defined as grade 0. When PComA is dilated but does not have additional anomalous extensive branches, this condition is identified as grade 1. Grade 2 is a dilated PComA accompanied by abnormally widespread branching. Grade 3 represents the absence of PComA visualization due to internal carotid artery occlusion.

Selecting the optimal model

To comprehensively evaluate predictive performance and ensure methodological rigor, this study conducted a comparative analysis across training, internal validation, and external validation cohorts using seven machine-learning models for variable selection: least absolute shrinkage and selection operator (LASSO) regression, K-nearest neighbors (KNN), support vector machine (SVM), decision tree, random forest, logistic regression, and Gaussian naive Bayes. These models were strategically selected to encompass diverse algorithmic paradigms, including linear vs. nonlinear modeling (LASSO vs. SVM), parametric vs. non-parametric approaches (logistic regression vs. KNN), and simplicity vs. ensemble complexity (decision tree vs. random forest). The LASSO model, known for its efficacy in feature selection and its ability to prevent overfitting in high-dimensional datasets, assumes a linear relationship between predictors and outcomes. For reproducibility, we set the regularization parameter (α) to 0.01 and used a maximum of 1000 iterations^[16]. KNN, which captures nonlinear relationships intuitively, was configured with k =5 neighbors and the Euclidean distance metric, although it can be computationally intensive and sensitive to the choice of k^[17]. SVM, particularly effective in handling nonlinear separability through kernel functions, was implemented with a Radial Basis Function kernel, where the penalty parameter C was set to 1.0 and gamma to “scale”^[18]. Decision Trees, offering simplicity and interpretability, were grown without pruning to a maximum depth of 5 to mitigate overfitting^[19]. Random Forests, which enhance prediction robustness by aggregating multiple decision trees, were set up with 100 estimators and a maximum depth of 10 to balance computational load and predictive performance^[20]. Logistic Regression, efficient and interpreted for linear relationships, was fitted with an L2 penalty and a regularization strength (C) of 1.0^[21]. Gaussian Naive Bayes, suitable for large datasets due to its simplicity and efficiency, assumes feature independence and was implemented with a variance smoothing parameter of 1e-9^[22]. The best model was chosen based on the C-index (concordance index), a measure that evaluates the degree of concordance between predicted and actual outcomes, especially in the context of survival analysis and other prediction tasks involving ordered outcomes. The C-index was instrumental in providing a robust criterion for model comparison, considering both discrimination ability and the capacity to handle censored data. This process helps to ensure the stability and generalizability of the selected model across diverse cohorts.

Predicting the preoperative bleeding risk

We employed the LASSO-logistic model to identify independent risk factors associated with preoperative bleeding in children with MMD. In our study, we applied a 10-fold cross-validation (nfolds = 10) approach to determine the optimal regularization parameter, λ. We explored 100 different λ values (nλ = 100) to effectively cover a wide range of regularization strengths. The final selection of variables was based on the λ value corresponding to lambda.1se, which was determined to be λ = 0.064. Additionally, for reproducibility, we set the random seed to 1234 to ensure consistent results across runs. This method initially screens preliminary variables using LASSO penalized regression before integrating the variables into logistic regression for modeling purposes. We applied LASSO regression using the glmnet package in R. We utilized cross-validation to determine the optimal lambda (λ) value, which controls the degree of regularization applied. This step is crucial as it helps to balance model fit and complexity. During the LASSO regression, coefficients for less important predictors are shrunk toward zero. The variables with nonzero coefficients after the regularization process were retained, indicating their significance in predicting the outcome. In contrast, traditional variable screening methods include univariate analysis and multivariable variable integration. However, univariate analysis is not recommended in the latest guidelines for PROBAST (PROBAST group Julius Center for Health Sciences and Primary Care, UMC Utrecht, Utrecht, the Netherlands) which assesses the quality, risk of bias, and applicability of prediction models. Eight potential risk factors other than the initial symptoms of preoperative bleeding were scrutinized and incorporated into the LASSO model to identify independent predictors of preoperative bleeding. The odds ratios (ORs) and 95% confidence intervals (CIs) for the risk of preoperative bleeding were computed using multivariate logistic regression analyses.

Construction of the nomogram

A predictive nomogram was developed to assess the risk of preoperative bleeding in pediatric patients with MMD. The performance of the nomogram was evaluated through three distinct methods: the concordance index (C-index) was utilized to measure the model’s discriminative ability, calibration was assessed using a calibration plot, and decision curve analysis (DCA) was employed to evaluate the clinical utility of the model. The C-index was calculated based on 1000 bootstrap resamples to determine the model’s internal consistency. Individual patient cumulative scores were derived from the nomogram. Additionally, receiver operating characteristic (ROC) curve analysis was performed to optimize Youden’s index and establish the most appropriate threshold value for prediction.

Statistical analysis

The data cohort was randomly divided into training and internal validation cohorts at a ratio of 7:3, and the variables were compared. Non-normal data are presented as median (interquartile range). In the univariate analysis, the chi-square or Fisher’s exact validation was used to analyze categorical variables. In contrast, the Student’s t-validation or rank-sum validation was used to examine continuous variables. In the training cohort, LASSO logistic regression was used for multivariate analysis to screen independent risk factors and build a prediction nomogram for the risk of preoperative bleeding. The Youden index was utilized to establish thresholds for the model’s two risk stratifications. Subsequently, we developed the nomogram model using the RSM package (version 6.6–0) and assessed its performance through ROC analysis, utilizing the pROC package (version 1.18.0), calibration curves (RSM package, version 6.6-0), and DCA with the ggDCA package (version 1.2). Statistical significance was defined as P < 0.05. R (R Foundation for Statistical Computing, Vienna, Austria; https://www.R-project.org/) and J MedCalc Statistical Software version 20.218 (MedCalc Software Ltd., Ostend, Belgium; https://www.medcalc.org; 2023) software packages were used for the analysis. We also used models with restricted cubic spline (RCS) to explore the association between age at onset and preoperative bleeding in pediatric patients with MMD.

Results

Patient characteristics

In total, 1350 pediatric patients with MMD were analyzed. Following PSM, 392 patients were included in the study, with 98 and 294 patients in the hemorrhagic and non-hemorrhagic MMD groups, respectively. A detailed flow diagram of patient selection is listed in Figure 1. The distribution of hemorrhagic MMD by age and location is detailed in Figure 2A,B. The selected demographic and clinical characteristics of the study population before and after PSM are summarized in Supplemental Digital Content 1, available at: http://links.lww.com/JS9/E361.

Figure 1.

Flowchart of patients through the study.

Figure 2.

Distribution of hemorrhagic cases by age and location. (A) Age at onset: The bar graph illustrates the distribution of hemorrhagic cases by age at onset among pediatric patients. (B) Location of hemorrhage: The bar graph displays the number of cases based on the location of hemorrhage.

Distribution of sex, initial symptoms, age at onset, family history, laterality, preoperative mRS score, AChA grade, PComA grade, and Suzuki stage was assessed in training, internal, and external validation cohorts. Among these cohorts, 137 out of 274 (50%), 61 out of 118 (51.7%), and 39 out of 70 (55.7%) individuals were male. In the training cohort consisting of 274 participants, the most common initial symptom was TIA, reported in 161 patients (58.8%). This was followed by cerebral infarction in 34 patients (12.4%), IVH in 23 patients (8.4%), and headache in 18 patients (6.6%). Other symptoms included syncope in six patients (2.2%) and ICH in nine patients (3.3%). A small proportion of patients were asymptomatic (4; 1.5%). In the internal validation cohort (N = 118), TIA occurred in 65 patients (55.1%), while cerebral infarction was reported in 21 patients (17.8%). The prevalence of IVH and headache were 10 cases (8.5%) and 9 cases (7.6%), respectively. Additionally, other initial symptoms included syncope in two patients (1.7%) and ICH in four patients (3.4%), whereas one patient (0.8%) was asymptomatic. For the external validation cohort comprising 70 participants, TIA was also the most prevalent initial symptom, reported in 41 patients (58.6%), followed by cerebral infarction in 13 patients (18.6%). This group also reported IVH in six patients (8.6%), headache in two patients (2.9%), syncope in two patients (2.9%), and ICH in two patients (2.9%). No patients were asymptomatic in this cohort. The training cohort, internal validation cohort, and external validation cohort had ages at onset of 7.9 ± 4.2, 8.5 ± 4.6, and 8.1 ± 4.5 years, respectively. Family history, laterality, preoperative mRS score, Suzuki stage, and AChA and PComA grades were compared between the cohorts. The distribution of sex, Suzuki stage, age at onset, family history, laterality, preoperative mRS score, AChA grade, and PComA grade did not significantly differ among the three cohorts (P > 0.05; Table 1).

Table 1.

Patient demographics and baseline characteristics in the training, internal validation cohort, and external validation cohort.

Characteristic	Cohort			P value
	Training cohort, N = 274	Internal validation cohort, N = 118	External validation cohort, N = 70
Sex				0.722
Male	138 (50.4%)	60 (50.8%)	39 (55.7%)
Female	136 (49.6%)	58 (49.2%)	31 (44.3%)
Initial symptoms
TIA	160 (58.4%)	66 (55.9%)	41 (58.6%)
Cerebral infarction	32 (11.7%)	23 (19.5%)	13 (18.6%)
SAH	2 (0.7%)	1 (0.8%)	2 (2.9%)
Headache	18 (6.6%)	9 (7.6%)	2 (2.9%)
Epilepsy	12 (4.4%)	1 (0.8%)	1 (1.4%)
IVH	23 (8.4%)	10 (8.5%)	6 (8.6%)
Syncope	7 (2.6%)	1 (0.8%)	2 (2.9%)
ICH + IVH	8 (2.9%)	1 (0.8%)	1 (1.4%)
ICH	11 (4.0%)	2 (1.7%)	2 (2.9%)
Asymptomatic	1 (0.4%)	4 (3.4%)	0 (0.0%)
Onset of age				0.988
Mean ± SD	8.1 ± 4.4	8.0 ± 4.2	8.1 ± 4.5
Family history				0.426
No	265 (96.7%)	111 (94.1%)	67 (95.7%)
Yes	9 (3.3%)	7 (5.9%)	3 (4.3%)
Laterality				0.824
Unilateral	14 (5.1%)	6 (5.1%)	2 (2.9%)
Bilateral	260 (94.9%)	112 (94.9%)	68 (97.1%)
Preoperative mRS				0.093
<2	215 (78.5%)	101 (85.6%)	51 (72.9%)
≥2	59 (21.5%)	17 (14.4%)	19 (27.1%)
AChA grade				0.322
0	90 (32.8%)	34 (28.8%)	24 (34.3%)
1	46 (16.8%)	19 (16.1%)	12 (17.1%)
2	49 (17.9%)	13 (11.0%)	8 (11.4%)
3	89 (32.5%)	52 (44.1%)	26 (37.1%)
PComA grade				0.616
0	75 (27.4%)	30 (25.4%)	20 (28.6%)
1	57 (20.8%)	23 (19.5%)	18 (25.7%)
2	55 (20.1%)	20 (16.9%)	8 (11.4%)
3	87 (31.8%)	45 (38.1%)	24 (34.3%)
Suzuki stage				0.023
<Ⅳ	245 (89.4%)	115 (97.5%)	62 (88.6%)
≥Ⅳ	29 (10.6%)	3 (2.5%)	8 (11.4%)

AChA, anterior choroidal artery; ICH, intraparenchymal hemorrhage; IVH, intraventricular hemorrhage; mRS, modified Rankin Scale; PComA, posterior communicating artery; SAH, subarachnoid hemorrhage.

Selecting the optimal model

This study indicates that LASSO model possessed higher mean C-index in the training cohort, internal validation cohort, and external validation cohort (corresponding to area under the curve [AUC], 91.5%, 78.4%, and 91.2%, mean C-index, 0.870) (Fig. 3). These findings suggest that LASSO regression is a superior variable screening method. Therefore, we will select the LASSO model for variable selection in subsequent analyses.

Figure 3.

A prediction model for preoperative bleeding was developed and validated using a comprehensive machine-learning approach. A total of seven distinct prediction models were evaluated, and the C-index for each model was calculated across all validation cohorts.

Nomogram variable screening

The original model incorporated several candidate predictors, including sex, age at onset, family history, laterality, preoperative mRS score, AChA grade, PComA grade, and Suzuki stage. Through LASSO regression analysis applied to the training cohort, the model was subsequently refined to include three potential predictors: age at onset, AChA grade, and PComA grade. The variance inflation factor values for the three variables are all below 10, suggesting the absence of multicollinearity and confirming that they are independent variables (Supplemental Digital Content Table S2, available at: http://links.lww.com/JS9/E362). Figure 4A illustrates a cross-validated error plot for the LASSO regression model, showcasing the error curve associated with the tuning parameter (λ), which was optimized at λ = 0.064. The variables selected by LASSO penalized regression in the training cohort are depicted in Figure 4B.

Figure 4.

Nomogram variable screening in training cohort. (A) LASSO regression model tuning with binomial deviance as the criterion. The plot shows the relationship between log(λ) and binomial deviance. (B) LASSO regression coefficient profile plot. Variables screened by LASSO penalized regression. The bar plot displays the estimated coefficients of the predictors selected at the λ value that minimizes the cross-validated error. LASSO, least absolute shrinkage and selection operator; PComA, posterior communicating artery; AChA, anterior choroidal artery.

Multivariate logistic regression

Further multivariate logistic analyses were carried out in the training cohorts to clarify whether the abovementioned three variables were independent risk factors for preoperative bleeding in pediatric patients with MMD. Finally, age at onset (OR = 1.17, 95% CI = 1.06–1.28, P = 0.001), AChA grade 1 (OR = 3.79, 95% CI = 1.02–14.14, P = 0.047) and grade 2 (OR = 14.98, 95% CI = 3.49–64.26, P < 0.001), as well as PComA grade 1 (OR = 8.44, 95% CI = 2.22–32.01, P = 0.002) and grade 2 (OR = 4.95, 95% CI = 1.15–21.39, P = 0.032), were identified as independent risk factors for preoperative bleeding in children (Fig. 5). The logistic regression model can be expressed as follows: $\log \left[\frac{\widehat{P}}{1-\widehat{P}}\right]$ = −3.927 + 0.154 (Onset of age) + 1.332 (AChA Grade 1) + 2.707 (AChA Grade 2) − 1.237 (AChA Grade 3) + 2.133 (PComA Grade 1) + 1.6 (PComA Grade 2) − 1.335 (PComA Grade 3).

Figure 5.

Forest plot of significant parameters in multivariate regression analysis. AChA, anterior choroidal artery; CI, confidence interval; OR, odds ratio; PComA, posterior communicating artery.

Development and implementation of a predictive nomogram using logistic regression for scoring equation extraction and risk stratification

The final logistic model included three independent predictors (age at onset, AChA grade, and PComA grade) and was developed as a simple-to-use nomogram (Fig. 6A). The total score was derived from the individual scores calculated using the nomogram. Patients in this study had total risk points of 0–280. We were >unable to determine the precise probabilities for the quantitative indicators due to axis scale limitations in the nomogram, thereby restricting its clinical applicability. We remedied this challenge by deriving polynomial equations for the predictors to ascertain the precise scores for each indicator and the respective risk for preoperative bleeding in pediatric patients with MMD associated with the total score. The risk for preoperative bleeding in pediatric patients with MMD was computed using the subsequent formula: Risk = −7.8 × 10⁻⁷ ×(ΣPoints)³ +0.00038 ×(ΣPoints)² − 0.0535 ×(ΣPoints) + 2.389, where ∑ Points represent the sum of points based on relevant clinical factors (Fig. 6B).

Figure 6.

Nomogram for predicting risk of preoperative bleeding in pediatric patients with MMD based on LASSO logistic regression. (A) Nomogram represents score assignment for each variable to calculate predicted probability of risk for preoperative bleeding in pediatric patients with MMD. Corresponding risk stratifications are labeled at bottom. (B) Risk calculation formula. Values of each variable correspond to different SCORE, and risk value obtained by summing scores and entering them into risk formula. (C‒E) ROC curve analysis of the three candidate risk indicators. AChA, anterior choroidal artery; AUC, area under time-dependent receiver operating characteristic (ROC) curve; PComA, posterior communicating artery.

The nomogram provides a quantified estimation of preoperative bleeding risk, effectively stratifying patients into three distinct categories: Low Risk, Middle Risk, and High Risk. Specifically, Low Risk corresponds to a bleeding probability of up to 0.1; Middle Risk ranges from a probability of 0.1–0.4; High Risk signifies a probability exceeding 0.4. This tool offers a clear and clinically valuable approach for evaluating bleeding risk, thereby facilitating individualized preoperative planning and informed decision-making. The nomogram’s systematic risk stratification facilitates clinical integration, with an interactive web application (https://lijingjie301.shinyapps.io/Hemorrhagic/) enabling real-time treatment guidance (Supplemental Digital Content Figure S1, available at: http://links.lww.com/JS9/E360).

To ensure the model was both regularized and parsimonious, we employed the LASSO method. LASSO is a regularization technique that performs variable selection by penalizing the absolute size of the regression coefficients, effectively shrinking some coefficients to zero. This helps to avoid overfitting and ensures that only the most relevant predictors are included in the final model. We used “10-fold cross-validation” to determine the optimal regularization parameter (λ) that minimized the cross-validated error. The λ value selected was the one that yielded a model within “one standard error of the minimum cross-validated error,” as recommended for achieving a balance between model complexity and predictive accuracy. The final model’s predictive value was further validated using ROC analysis, which yielded AUC values of “0.711 (age at onset), 0.836 (AChA grade), and 0.839 (PComA grade),” indicating good discriminative ability (Fig. 6C–E).

C-Index, calibration, and clinical utility of a nomogram

The model demonstrated favorable discrimination with an AUC of 0.916 (95% CI: 0.871–0.961) in the training cohort (Fig. 7A). The discrimination of the model was confirmed to be satisfactory using the internal validation cohort, with an AUC of 0.842 (95% CI: 0.742–0.941; Fig. 7B). Furthermore, in the external validation cohort, the model maintained a satisfactory discrimination with an AUC of 0.892 (95% CI: 0.798–0.986; Fig. 7C). The calibration curve indicated a strong correlation between the model’s predictions and the actual observations in the training cohort (Fig. 7D). The Hosmer–Lemeshow test yielded a nonsignificant P = 0.503, indicating that the predicted and actual probabilities were highly consistent. Similarly, in the internal validation cohort, we observed good calibration, as indicated by a nonsignificant p-value obtained from the Hosmer–Lemeshow test (Fig. 7E). The calibration curve also demonstrated good agreement between predicted and observed probabilities (Fig. 7F). The DCA demonstrated that using the model was more beneficial compared to treating all patients or no patients in the training cohort when the threshold probabilities for clinicians or patients fell within the range of 10%–90% (Fig. 7G). The DCA indicated that in the internal validation cohort, the model exhibited a higher net benefit when the threshold probabilities were within the range of 10%–75% (Fig. 7H). The DCA for the external validation cohort showed that the model was beneficial when threshold probabilities were between 10% and 75% (Fig. 7I). These results confirm the model’s robustness and utility across different datasets for clinical decision-making.

Figure 7.

C-index, calibration, and application of nomogram to predict risk of preoperative bleeding in pediatric patients with MMD. C-index (A‒C), calibration (D‒F), and decision curve (G‒I) analyses of diagnostic nomogram in training, internal, and external validation cohorts. AUC, area under the receiver operator characteristic curve.

Association between the age at onset and preoperative bleeding

This study determined that age at onset independently predicted preoperative bleeding. Therefore, we used the RCS function to explore the correlation between age and preoperative bleeding. Possible nonlinear relationships between the age at onset and preoperative bleeding were examined using a logistic regression model with RCS (Fig. 8). The inflection point (age: 8 years) was cohort as the cut-off value. In pediatric MMD, the risk of preoperative bleeding was greater when the patient was aged >8 years.

Figure 8.

Association between age at onset and preoperative bleeding.

Discussion

Pediatric hemorrhagic MMD is rare, with a 3% incidence in children versus 25%–60% in adults in Korea, with little clinical evidence of its risk for preoperative bleeding^[23,24]. Previous studies, including Liu et al (n = 30)^[8], Ge et al (n = 46)^[25], and Ahn et al (n = 13)^[5], were limited by small cohorts. Our multi-institutional study, the largest to date, included 392 patients (training: N = 274; internal validation: N = 118) and an independent external cohort (n = 70). This aggregate (N = 462) adheres to rare disease research standards (cohorts ≥50–100 for validation). The external cohort demonstrated strong discriminative performance (AUC = 0.892, 95% CI: 0.798–0.986), with tighter CIs than prior studies. While larger samples remain ideal, our study establishes a new benchmark in scale and methodological rigor for pediatric hemorrhagic MMD research.

Accurate preoperative risk stratification for hemorrhage in pediatric MMD is critical to guide proactive management and optimize surgical outcomes. In this study, LASSO-logistic regression identified three independent predictors of bleeding risk: older age at symptom onset, AChA dilation (grades 1–2), and PComA dilation (grades 1–2). These variables were integrated into a clinically applicable nomogram, validated across multi-institutional cohorts, which quantifies individualized bleeding probabilities using a parsimonious set of predictors. RCS analysis further delineated a nonlinear relationship between age and hemorrhage risk, with a marked escalation beyond 8 years of age. The model’s discriminative performance (AUC: 0.916 training, 0.892 external validation) and calibration accuracy (Hosmer–Lemeshow P = 0.503) underscore its potential to enhance preoperative decision-making, particularly in identifying high-risk children who may benefit from intensified monitoring or early intervention. By balancing simplicity with precision, this tool addresses a critical gap in pediatric MMD care, where evidence-based risk stratification has historically been limited.

Prior studies in adults with MMD have identified hemorrhage risk factors, including age at onset, choroidal anastomosis, hypertension^[26], and cerebral microbleeds^[27]. Imaging assessments in adults further established choroidal anastomoses^[28] as sources of posterior circulation hemorrhage and independent predictors of rebleeding risk^[29]. In contrast, pediatric data remain limited to small cohorts linking AChA and PComA dilatation to hemorrhagic events^[8]. Our findings corroborate these associations while identifying age >8 years as a novel inflection point for escalating preoperative bleeding risk. Notably, hemorrhagic presentation may occur as the initial manifestation in pediatric MMD, underscoring the clinical urgency of early risk stratification. By integrating these validated predictors into a pragmatic nomogram, our model addresses a critical evidence gap in pediatric neurosurgical practice.

Age at onset emerged as an independent predictor of preoperative bleeding in pediatric MMD, though its role remains debated across age groups. Prior adult MMD studies report decreasing hemorrhage risk with advancing age^[26], contrasting with our pediatric findings where older onset age (>8 years) predicted higher preoperative bleeding risk. This divergence may reflect distinct disease pathophysiology between pediatric and adult populations. Our RCS analysis established age 8 as a critical threshold for bleeding risk elevation, corroborating our earlier observation of older onset ages in hemorrhagic versus ischemic pediatric MMD cases (12.6 ± 3.18 vs. 9.4 ± 3.8 years; P < 0.05)^[24]. While some adult studies conflict regarding age’s association with hemorrhage risk^[30,31], our findings demonstrate robust correlations in pediatric populations, validated across internal and external cohorts. This study advances clinical understanding by quantitatively defining age-related risk thresholds through nonlinear modeling, providing actionable thresholds for pediatric surgical decision-making.

The age-dependent hemorrhage risk in MMD reflects distinct pathophysiological mechanisms across developmental stages. Younger children predominantly experience ischemic strokes from rapid ICA stenosis and insufficient collaterals, compounded by sparse, unstable moyamoya vessels vulnerable to watershed infarctions during metabolic stress^[32], potentially influenced by genomic factors^[33]. In contrast, patients >8 years demonstrate elevated preoperative bleeding risk, potentially involving age-related collateral network maturation that increases hemodynamic stress on fragile vessels^[34], with heightened physical activity exacerbating vascular vulnerability. This pediatric pattern diverges from adult MMD where hemorrhagic risk peaks nonlinearly between ages 46 and 55 years^[35] before declining in elderly patients – a phenomenon possibly attributed to stabilized collaterals compensating for reduced flow, lower cerebral metabolic demands^[36], and improved blood pressure control through antihypertensive therapies^[37]. These contrasting age-specific risk profiles underscore the necessity for tailored management strategies addressing pediatric hemodynamic fragility while leveraging adult vascular stabilization mechanisms.

AChA dilation, an established hemorrhagic risk factor in adult MMD^[26,38,39], demonstrates similar prognostic significance in pediatric populations. While only one pediatric study previously linked AChA dilation to hemorrhagic presentation^[8], our findings corroborate this association through quantitative grading analysis – a critical advancement beyond adult-focused studies that lacked stratification. Multivariate analysis identified AChA grades 1–2 as independent predictors, offering clinically actionable thresholds for preoperative risk assessment. This granular stratification enables targeted intervention strategies for high-risk pediatric patients, contrasting with adult literature that primarily associates choroidal collaterals with rebleeding risk without quantitative grading^[40]. The operational value of this grading system lies in its capacity to guide proactive management, potentially mitigating long-term hemorrhagic complications through early identification of vulnerable vascular phenotypes.

PComA dilation, a validated hemorrhagic risk marker in adult MMD^[38,41], extends to pediatric populations in our analysis, though with no ethnic variations – its predictive power remains confined to Asian cohorts despite lacking confirmation in European populations^[42]. These racial disparities in collateral pathway significance may reflect underlying differences in MMD pathophysiology. While adult studies associate posterior circulation involvement with initial hemorrhage events^[15], pediatric data remain exceptionally limited, with our prior small-sample investigation standing as the sole preceding pediatric reference^[8]. Multivariate analysis established PComA grades 1–2 as independent predictors, demonstrating escalating preoperative bleeding risk at higher grades. The nomogram further revealed PComA dilation’s superior predictive value over isolated AChA involvement, providing critical operative implications for prioritizing vascular monitoring in pediatric cases with posterior circulation anomalies. This grading system’s clinical utility lies in its capacity to stratify surgical risk while addressing the historical neglect of pediatric-specific collateral assessment in MMD management.

Our study introduces the first validated nomogram for preoperative bleeding risk assessment in pediatric MMD, integrating three accessible clinical variables (age at onset, AChA grades 1–2, and PComA grades 1–2) into a practical predictive tool. Derived from the largest global pediatric hemorrhagic MMD cohort analyzed to date, this model demonstrated robust performance through rigorous Multivariate Predictive Model-compliant development and validation^[43], including bootstrap-corrected C-index and AUC with optimal calibration. The nomogram’s clinical strength lies in its simplicity and reproducibility – validated across internal and external cohorts – enabling rapid risk stratification using routinely available preoperative data. By quantifying posterior circulation collateral grades rather than binary classifications, this tool advances pediatric-specific risk prediction beyond adult-centric models, offering neurosurgeons an evidence-based framework for perioperative decision-making in this vulnerable population.

The nomogram’s simplicity (age at onset, AChA/PComA grades 1–2) facilitates preoperative risk stratification, enabling tailored interventions like prioritized surgery for high-risk patients (scores >168) with strict monitoring, while moderate-risk cases may require perfusion studies to optimize collateral evaluation. This pragmatic tool supports clinical decision-making by aligning with precision neurosurgical approaches through individualized risk profiling, particularly in guiding revascularization strategies and hemodynamic management for pediatric MMD.

While our nomogram demonstrated robust predictive accuracy in the current cohort, several limitations warrant careful consideration. First, although this study represents the largest pediatric hemorrhagic MMD cohort analyzed to date, the nomogram’s generalizability requires further validation in ethnically and geographically diverse populations to account for potential variations in disease pathophysiology and collateral vessel patterns. Notably, East Asian cohorts exhibit distinct genetic profiles (RNF213 p.R4810K mutation prevalence)^[4] and phenotypic trajectories compared to European/African populations, where MMD is more strongly associated with autoimmune comorbidities and posterior circulation involvement^[44]. These disparities may alter bleeding risk predictors, as collateral vessel stability and hemodynamic stress responses are modulated by both genetic and environmental factors (e.g. regional variations in dietary nitrates and air pollution exposure). Second, while PSM balanced observed confounders (sex, Suzuki stage, preoperative mRS), residual confounding from unmeasured variables – including medication adherence and epigenetic factors – may persist despite multivariate adjustments, highlighting inherent limitations of retrospective analyses that necessitate prospective validation with detailed pharmacologic and comorbidity documentation. Third, the retrospective design inherently risks selection bias, particularly regarding excluded cases with missing data. Although multiple imputation was employed, non-random missingness may distort risk estimates. Finally, inter-institutional heterogeneity in diagnostic protocols could affect model performance across settings. Future research should prioritize multicenter validation across diverse populations, multi-omic profiling to uncover ethnicity-specific risk, longitudinal pediatric studies assessing hemodynamic and genetic impacts, and AI-driven dynamic risk modeling integrating real-time biomarkers and integrated machine learning to enable personalized, globally applicable pediatric MMD management.

Conclusion

The nomogram demonstrated high accuracy in predicting the risk of preoperative bleeding in pediatric patients with MMD. This predictive accuracy may enhance preoperative evaluation by surgeons, leading to more proactive treatment and stricter monitoring of patients with a high risk of preoperative bleeding.

Ethical approval

The institutional review board and ethics committee of the Fifth Medical Centre of Chinese PLA General Hospital approved the study (Approval number: ky-2020-9-22), and due to the retrospective of the data, an exemption from obtaining informed consent was granted.

Consent

Due to the retrospective of the data, an exemption from obtaining informed consent was granted.

Sources of funding

This study was supported by grants from the National Natural Science Foundation of China (grant numbers 82171280, 82201451, and 82172021).

Author contributions

L.D., X.B., and C.H. designed the study. Q.G. and M.X. wrote the manuscript and performed the medical data analysis. Q.-N.W. contributed to the manuscript discussion, figures, and supplementary material. All the authors contributed to the manuscript and approved the submitted version.

Conflicts of interest disclosure

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Research registration unique identifying number (UIN)

This study has been registered in the Chinese Clinical trial registry (registration number: ChiCTR2200064160).

Guarantor

Lian Duan.

Provenance and peer review

Not commissioned, externally peer-reviewed.

Data availability statement

Original data were generated and stored at the Fifth Medical Center of the People’s Liberation Army of China (PLA) General Hospital. Data supporting these results may be obtained from the corresponding authors if the requirements are reasonable.