We conducted a retrospective cohort study utilizing data from two centers: The Second Affiliated Hospital of Wenzhou Medical University (Derivation cohort) and The First Affiliated Hospital of Wenzhou Medical University (Validation cohort). The study period for the derivation cohort spanned from 1 June 2018 to 28 February 2024, while the validation cohort data was collected from 1 August 2021 to 28 February 2024. Participants were patients aged ≥18 years with hypertensive ICH (HICH). HICH was defined by the presence of ICH with a history of hypertension. Exclusion criteria included: patients who underwent surgical treatment before the repeat CT scan; hemorrhage secondary to cerebral arteriovenous malformations, trauma, intracranial aneurysms, brain tumors, or hemorrhagic transformation of brain infarction; coagulopathy-related brain hemorrhage, and no repeat head CT scan within 24 h after the initial CT scan. This study was approved by the Ethics Committee of the Second Affiliated Hospital of Wenzhou Medical University. The informed consent was waived due to the retrospective nature.

We identified patients through a comprehensive electronic medical records system at both centers. The search criteria included relevant International Classification of Diseases (ICD) codes for HICH, as well as key terms associated with HICH. The medical records of potential participants were manually reviewed by two researchers to ensure that all inclusion and exclusion criteria were met.

To ensure the robustness of our predictive model, we adhered to the principle of Events Per Variable (EPV), which is a widely accepted criterion for determining the adequate sample size in logistic regression analysis (11). The EPV principle recommends that there should be at least 10 events (or cases) per predictor variable to avoid overfitting and to ensure the stability of the regression coefficients. Considering an expected event rate of approximately 20–30% and allowing for a dropout rate of 10%, we estimated that we would need a total sample size of at least 500 patients in the modeling cohort to observe approximately 100 events. For the validation cohort, we aimed to include at least 200 patients to validate the model derived from the modeling cohort, ensuring that the model’s predictive performance could be generalized to an independent set of patients.

Candidate variables included demographics, medical history, clinical presentation, laboratory results, and imaging findings. Specific variables were selected based on their potential association with HE, as informed by previous literature and clinical expertise. Demographic (Age, Gender), vital signs upon Emergency Department (ED) arrival, GCS, anticoagulant/antiplatelet therapy, time from symptom onset to head CT scan, baseline hematoma volume (calculated using the ABC/2 method or volumetric analysis), hematoma shape (irregularity, lobulation), and non-homogeneous density which was manifesting as any of the following signs: blend sign, hypodensities, black hole sign, swirl sign, island sign, or liquification, biochemical indicators (white blood cell count, hemoglobin, platelet count, serum glucose, serum creatinine, international normalized ratio (INR)). Hypodensities, defined as areas of lower attenuation; liquification: presence of fluid levels within the hematoma; Island sign was defined as ≥3 separate small hematomas dispersed and distinct from the main hematoma; or ≥4 small hematomas, several of which can be connected to the main hematoma. Black hole sign was characterized as a well-demarcated hypodense area within the hyperdense hematoma, which may be round, oval, or strip-shaped, and does not connect with the surrounding brain tissue, with a CT value difference of at least 28 Hounsfield Unit (HU) from the surrounding hematoma. Swirl sign was described as a hypodense or isodense area within the hyperdense hematoma, with variable shapes such as round, strip-like, or irregular. Blend sign: Refers to the coexistence of relatively hyperdense and hypodense areas within the hematoma, with a clear boundary between them, and a CT value difference of more than 18 HU (10). Uncontrolled blood pressure was defined as systolic blood pressure (SBP) ≥ 140 mmHg recorded within 1 h of hospital admission.

The primary outcome was the occurrence of HE within 24 h of the onset of symptoms, defined as an increase in hematoma volume ≥33% or ≥12.5 mL on follow-up imaging (5).

Continuous variables were presented as mean ± standard deviation (SD) or median with interquartile range (IQR), depending on the distribution. To compare baseline characteristics between two cohorts, the student’s t-test or Mann–Whitney U test was used for continuous variables, as appropriate. For categorical variables, the chi-square test or Fisher’s exact test was applied.

For the derivation of the predictive model for HE in patients with HICH, we employed a rigorous statistical approach. Candidate variables were selected based on two criteria: (1) their established link to HE as reported in existing literature, and (2) their availability in routine hospital settings without requiring additional intervention for assessment. This variable selection ensured that all variables included in the final model would be clinically relevant. We utilized a standard logistic regression model with HE (yes/no) as the dependent variable. All preselected variables were simultaneously entered into the model. Variables were retained in the final model only those with p value<0.05 from the regression analysis. We aimed to construct a predictive scoring system based on the final model. The scoring system was developed by dividing the regression coefficient of each predictor variable by the smallest coefficient from the model, thereby assigning a whole number or half-point score to each variable.

The new score’s discrimination for HE was examined by the area under the receiver operating characteristic curve (AUC) and calibration was conducted using calibration curves and the Hosmer and Lemeshow (H–L) test. The DeLong’s non-parametric method was used to compare the AUC values. To assess the clinical utility of the new score in predicting HE, Decision Curve Analysis (DCA) was conducted to compare the net benefits of the new score, BRAIN, PREDICT and the 9-point across various threshold probabilities.

Missing data were predominantly observed in laboratory indicators, with the rate of missingness being less 5% for each indicator. Missing data filling is carried out by using the multiple imputation. Since our prediction model is not intended to incorporate laboratory indicators, missing data has no impact on our model construction. All analyses were conducted using R, version 4.4.0 (R Foundation for Statistical Computing, Vienna, Austria), and a p value < 0.05 was considered to be statistically significant.

In our analysis, the derivation cohort encompassed 567 patients, and the validation cohort comprised 378 patients (Figure 1). While the baseline characteristics of the two cohorts were largely similar, there were notable differences in vital signs upon admission and imaging features, specifically hematoma volume and density characteristics. Despite these differences, the incidence of HE was comparable, occurring in 23.8% of the derivation cohort and 22% of the validation cohort. A detailed comparison of the baseline characteristics of both cohorts is presented in Table 1.

The multivariate analysis revealed seven independent predictors of HE: Age, uncontrolled blood pressure, hematoma volume, irregularity/lobulation shape, non-homogeneous density, CT scan performed ≤6 h after symptom onset, and the use of anticoagulation/antiplatelet therapy (Table 2). These variables were chosen based on their established link to HE as reported in existing literature and their availability in routine hospital settings.

We developed a novel scoring system specifically designed to predict HE in patients with HICH. We named this scoring system the HE-VSD-A₂TP score, which is derived from the seven variables it encompasses: Hematoma Volume, Shape, Density, Age, Anticoagulation/antiplatelet, within 6 h from symptom onset to CT, uncontrolled blood Pressure. Each variable was assigned a score proportional to its regression coefficient as determined from the multivariate logistic regression analysis (Table 2). The score for each predictor variable was calculated by dividing the regression coefficient of the variable by the smallest regression coefficient among the significant predictors, ensuring that all scores were on a comparable scale. This process allowed for the transformation of the regression coefficients into clinically interpretable scores. The HE-VSD-A₂TP score for an individual patient was calculated by summing the scores of each predictor variable that applied to that patient. This cumulative score provided a quantitative estimate of the patient’s risk for HE. The total score ranged from 0 to a maximum of 12.5, with a higher score indicating a greater risk of hematoma expansion.

We compared the performance of the HE-VSD-A₂TP score with established models, PREDICT, 9 Point and BRAIN score, in predicting HE in patients with HICH. In the derivation cohort, the AUROC for the HE-VSD-A₂TP score was 0.871 (95% CI, 0.839–0.904), which was significantly higher than that of the PREDICT score at 0.757 (95% CI, 0.711–0.802), the 9 Point score at 0.654 (95% CI, 0.601–0.706), and the BRAIN score at 0.739 (95% CI, 0.695–0.782) (Figure 2A and Table 3). This indicates that the HE-VSD-A₂TP score demonstrated superior discrimination in predicting HE compared to the other scoring systems. In the validation cohort, the AUROC values for the HE-VSD-A₂TP score remained high at 0.858 (95% CI, 0.819–0.897) and were significantly higher than those for the PREDICT score at 0.584 (95% CI, 0.510–0.657), the 9 Point score at 0.711 (95% CI, 0.644–0.779), and the BRAIN score at 0.762 (95% CI, 0.707–0.818) (Figure 2B and Table 3). The consistency of the HE-VSD-A₂TP score’s performance across both cohorts highlights its robustness and generalizability in predicting HE. Furthermore, the HE-VSD-A₂TP score outperformed the PREDICT, 9 Point, and BRAIN scores in terms of sensitivity, specificity, LHR, NPV, and PPV. The detailed performance metrics for each score are presented in Table 4. The superior performance of the HE-VSD-A₂TP score across these metrics underscores its clinical utility in more accurately identifying patients at risk of HE.

In addition to the discrimination analysis, we also evaluated the calibration of the HE-VSD-A2TP score compared to the PREDICT, 9 Point, and BRAIN scores by H–L test and calibration curves (Figure 3 and Table 3). In the derivation cohort, the HE-VSD-A₂TP score demonstrated excellent calibration with a H-L Chi-square value of 7.569 and a p-value of 0.271. The deviation correction curve generated by the bootstrap method is displayed very close to the reference line, indicating that the HE-VSD-A2TP prediction of HE occurrence is still highly consistent with the actual HE occurrence. In contrast, the PREDICT score, 9 Point score and BRAIN score exhibited the poor calibration (Figure 3A and Table 3). In the validation cohort, the calibration performance of the HE-VSD-A₂TP score was consistent with that observed in the derivation cohort (Figure 3B and Table 3).

The stratification of HE risk in HICH patients was determined using the HE-VSD-A₂TP score, which was calculated by summing the points assigned to each identified risk factor. The risk stratification was categorized into three groups: low risk (0–4.5 points), moderate risk (5–7 points), and high risk (7.5–12.5 points). The HE-VSD-A₂TP score effectively categorized patients into low-risk, moderate-risk, and high-risk groups for HE in both the derivation and validation cohorts. Each risk group had a predicted HE rate that closely mirrored the actual HE rates (Table 5).

Decision curve analysis (DCA) was conducted to evaluate the clinical utility of the HE-VSD-A₂TP score compared to the PREDICT, 9 Point, and BRAIN scores in both the derivation and validation cohorts (Figure 4). The DCA graphically represents the net benefit of using each scoring system across a range of threshold probabilities for predicting HE. The DCA results from both cohorts consistently show that the HE-VSD-A₂TP score outperforms the PREDICT, 9 Point, and BRAIN scores in predicting HE. This suggests that the HE-VSD-A₂TP score is more likely to identify patients who would benefit from targeted interventions to prevent HE, thereby potentially improving patient outcomes.