Our study enrolled 410 patients with aSAH admitted to the Neurosurgery Department of the Affiliated People’s Hospital of Jiangsu University between January 2020 and December 2023. Data were retrospectively collected using a clinical research data platform. Patients were excluded if they did not have aneurysms, had inadequate imaging quality, were hospitalized for fewer than 5 days, refused participation, died within 30 days of treatment, had other brain diseases (e.g., vascular malformations, brain tumors), or had a history of hydrocephalus or shunt surgery. Diagnostic Criteria for Chronic Hydrocephalus: 1. Time criterion: Defined as 14 days or more after subarachnoid hemorrhage (SAH). 2. Imaging criteria: Meeting at least one of the following: Evans index >0.30, callosal angle <90°, or unilateral/bilateral temporal horn width >2 mm. 3. Exclusion criteria: Congenital hydrocephalus, cerebral atrophy, brain tumors, or other causes of ventricular enlargement were excluded. Hospitalization duration <5 days (82 cases).

Death within 30 days (20 cases); History of cranial surgery (26 cases); Missing or poor-quality imaging data (71 cases); Presence of other intracranial diseases (8 cases); History of pre-existing hydrocephalus or shunt surgery (23 cases). After rigorous screening and data organization, medical records from 180 patients were included in the statistical analysis and randomly assigned to a training cohort and a validation cohort in a 7:3 ratio (Figure 1). This retrospective study was approved by the Medical Ethics Committee of the Affiliated People’s Hospital of Jiangsu University (Approval Number: K-202400164 W) and conducted in adherence to the principles of the Helsinki Declaration. As all patient data were anonymized and de-identified prior to analysis, informed consent was not required.

This study meticulously extracted demographic and clinical data from the medical record system. In addition to recording patients’ basic information, we assessed comorbidities, including hypertension [as described in our previous study (8)], diabetes, heart disease, stroke, and prior history of surgery. Key clinical indicators, such as the Glasgow Coma Scale (GCS) score, Hunt and Hess grade, World Federation of Neurosurgical Societies (WFNS) grade, and modified Fisher grade at admission, were also documented. The diagnosis of delayed cerebral ischemia (DCI) was based on established field standards, incorporating clinical symptoms (new-onset neurological deficits, impaired consciousness) combined with imaging evidence of ischemia (new infarct foci indicated by CT). We organized examination data to include the location of aneurysms, surgical procedures, and the presence of craniectomy and ventricular hemorrhage for each patient. The blood clearance estimated are as follows: The timeline starts from hospital admission (or symptom onset) and continues until follow-up CT scans show no significant high-density blood signals in the subarachnoid space and ventricular system (as independently assessed by two radiologists, with disagreements resolved by a third evaluator). The first day on which “complete absorption” is observed is recorded as the “clearance interval.”

Postoperative data included details about lateral ventricular drainage, ventriculoperitoneal (V-P) shunt, and lumbar puncture drainage. Complications were also recorded, such as acute or subacute hydrocephalus, cerebral hernia, and delayed cerebral ischemia and the time required for the complete clearance of intracranial hematomas. All data were rigorously reviewed by clinicians to ensure accuracy and consistency. Within the first week of admission, cerebrospinal fluid (CSF) tests were performed multiple times to measure chloride, glucose, cell count, and protein levels. Measurement was taken within 7 days after admission, using the results from the first lumbar puncture (LP) or external ventricular drainage (EVD) specimen. The 7-day window was chosen because it corresponds to the critical phase of arachnoid inflammation and CSF circulation abnormalities triggered by blood breakdown products. For serological indicators, the reference ranges were: white blood cell count 3.5–9.5 × 10⁹/L and C-reactive protein (CRP) 0–5 mg/L. Electrolyte levels were also monitored, with reference ranges as follows: calcium (Ca) 2.08–2.66 mmol/L, potassium (K) 3.5–5.1 mmol/L, sodium (Na) 135–145 mmol/L, and chloride (Cl) 95–106 mmol/L. Measurement was taken within 24 h after admission, using the first test result as the analytical value. If multiple tests were performed within 24 h, the earliest baseline data were selected. Radiological features were obtained by manually measuring the initial CT scan images upon admission, as described in our previous study (8).

Due to the relatively concealed formation process of hydrocephalus, it is difficult to detect it through a single CT scan. As noted in our previous study, white matter surrounding the ventricles changed after 7 days post-SAH (8). For the calculation of radscore, CT scans from the 7th day of admission were selected. The region of interest (ROI) was defined as the periventricular white matter within a 1 cm diameter around the anterior horns of the lateral ventricles. CT images in DICOM format were imported into 3D Slicer (version 5.2.2), and the delineation of ROI was performed by experienced resident physicians and validated by a senior radiologist with over 20 years of experience in diagnosing brain diseases. Radiomic features were extracted from the periventricular white matter around the anterior horns of both lateral ventricles using Pyradiomics (25), yielding 1,395 features per ROI after excluding morphological features. The features extracted from both the left and right sides were collectively compared between the two patient groups using the Mann–Whitney U test, and those with p ≥ 0.05 were excluded. Least absolute shrinkage and selection operator (LASSO) regression, with 10-fold cross-validation, identified the optimal regularization parameter (λ), an intercept (α), and the corresponding coefficients (β) for each feature. Seven features and their corresponding coefficients were retained for calculating the final radiomics score (Figure 2).

The 3D-Unet automatic segmentation model (Figure 3) was trained using initial CT scan images obtained at admission. The images were standardized and resampled to a uniform spatial resolution of 0.5 × 0.5 × 5 mm. After removing the skull, hemorrhagic lesions were annotated slice by slice. The dataset was split into training, validation, and testing sets in a 7:1:2 ratio. Data diversity was enhanced by horizontal flipping of the training dataset. Detailed parameters of the radiomics and segmentation models are provided in Supplementary material.

Univariate logistic regression (LR) was employed to analyze clinical and radiological parameters. Significant attributes (p < 0.05) were further analyzed using LASSO regression with 10-fold cross-validation to refine the selection. Multivariate LR identified significant risk factors for chronic hydrocephalus after aSAH (13). These factors were integrated into predictive models using both LR and ML algorithms, including extreme gradient boosting (XGB), random forest (RF), support vector machine (SVM), decision tree (DT), and k-nearest neighbor (KNN). A user-friendly nomogram was developed for healthcare professionals (26, 27).

Statistical analyses were conducted using SPSS 22.0 and R software (version 4.1) (8, 13). Continuous variables were presented as median [interquartile ranges (IQR)] or mean ± standard deviation (SD), depending on the results of the Shapiro–Wilk normality test. Categorical data were expressed as proportions. Chi-square tests, Fisher’s exact tests, and Mann–Whitney U tests were used to assess differences among cohorts. Univariate LR analysis was conducted to identify potential risk factors associated with the study outcomes, with variables showing a p < 0.05 included in a stepwise multivariable LR model. The nomogram’s predictive performance was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC), comparing predicted probabilities to actual outcomes. Calibration curves were used to assess model fidelity, and goodness-of-fit testing was performed, with calibration performance measured using Boullier scores. Clinical utility of the refined model was further evaluated using decision curve analysis (DCA) across different probability thresholds for both groups. ML algorithms, renowned for their superior predictive performance, have demonstrated superior outcomes compared to traditional regression methods for large datasets (28). To enhance predictive accuracy, six ML algorithms were applied: XGB, LR, RF, SVM, deep regression (DR), and KNN. For model training, 70% of the dataset was randomly selected, while the remaining 30% was reserved for testing. To mitigate overfitting, appropriate adjustments were made during training, and optimal hyperparameters were determined using five-fold cross-validation. The optimized model was implemented in R to predict the risk of chronic hydrocephalus following aSAH. Statistical significance was defined as a p < 0.05.

A retrospective analysis was conducted using data from 410 patients enrolled at our institution, of whom 180 were included in this study. The cohort was randomly divided into training and testing groups in a 7:3 ratio. No significant differences were found across all variables between the training and testing groups (Table 1) (p > 0.05). The overall incidence of chronic hydrocephalus was 22.2% (40/180), with an incidence of 24.8% (31/125) in the training group and 16.3% (9/55) in the testing group.

The final seven selected radiomic features and their corresponding coefficients, were used to calculate the radiomic score (Supplementary Figure 1; Table 2).

Waterfall plots and box plots were used to show the distribution of the radscore in both the training and test cohorts (Figure 4).

These findings highlighted the consistent performance between the two cohorts and effectively demonstrated how the radiomic features differentiated between outcome groups.

The training process and internal validation results of the 3D-Unet automatic segmentation model are shown in Supplementary Figure 2. Following training, 36 images were used as the test set to quantitatively evaluate the segmentation performance of the model using several evaluation metrics: Dice similarity coefficient (DSC), Intersection over Union (IoU), Hausdorff Distance (HD), and Average Symmetric Surface Distance (ASSD). The performance metrics for the model are presented in Supplementary Table 1.

A forest plot was generated to visually illustrate the various risk factors associated with chronic hydrocephalus after aSAH using univariate LR analysis (Figure 5).

This analysis identified 21 risk factors with p < 0.05, and LASSO regression was used to determine nine prominent factors associated with chronic hydrocephalus after aSAH (Figure 6).

Ultimately, six key determinants were identified: cerebrospinal fluid lactic acid level, Na, corpus callosum angle, interval to blood clearance, periventricular white matter changes, and hematoma volume. These factors were further analyzed using a combination of ML algorithms, including LR, XGB, LGBM, RF, SVM, DT, and KNN. In the training set, KNN, RF, and XGB models exhibited the highest performance, achieving AUC values of 1.00, 1.00, and 0.96, respectively, indicating exceptional predictive capabilities. Other models, such as LR and SVM, also performed well.

However, in the test set, the performance of some models declined. KNN and RF, in particular, appeared to suffer from overfitting, making them unsuitable for further evaluation in subsequent model-building processes. The SVM and LR models demonstrated greater stability, with minimal variation in performance across different datasets.

The LGBM model showed poor performance in both the training and test sets, struggling to distinguish between samples.

Therefore, when selecting a model for predicting chronic hydrocephalus, it is crucial to account for potential overfitting. While some models performed exceptionally well in the training set, LR, SVM, NB, and DT demonstrated more robust performance with better generalization. LR outperformed other models in the test set, making it the preferred choice for predictive tasks (Figure 7).

The calibration curves revealed that both the LR and SVM models showed relatively stable performance across both the training and test datasets. The slight variation between their predicted and actual probabilities highlighted their strong generalization capabilities. In contrast, while the DT and XGB models performed well in the training dataset, their performance significantly deteriorated in the test dataset, suggesting a significant risk of overfitting (Figure 8).

The Precision-Recall (P-R) curve (Figure 9) demonstrated a balance between precision (positive predictive value) and recall (sensitivity), which might be particularly valuable in evaluating classifiers in imbalanced class situations.

Furthermore, the calibration curves for the LR model in both the training and test sets showed optimal performance (Figure 10), indicating strong consistency between the models’ predictions and actual outcomes.

Based on these analyses, we developed a nomogram to predict the likelihood of chronic hydrocephalus after aSAH (Figure 11).

DCA further highlighted the model’s advantage in predicting chronic hydrocephalus when the threshold probability ranged from 7 to 47% in the test set (Figure 12).

The clinical impact curve (CIC) demonstrated the consistency of the model in predicting the number of high-risk patients for hydrocephalus compared to those actually diagnosed with the condition (Figure 13).

Patients with aSAH commonly experience electrolyte disturbances during hospitalization. This study identified a significant association between hypernatremia and the development of chronic hydrocephalus. Hypernatremia frequently manifests early post-SAH (30, 31), and may result from hypothalamic dysfunction, which impairs arginine vasopressin secretion. This condition is characterized by elevated extracellular sodium levels, leading to hyperosmolarity, with approximately 90% of the body’s sodium ions localized in the extracellular space (32). In a hyperosmolar state, water molecules within the cells move to the extracellular space to maintain osmotic balance, which reduces cell volume and leads to the contraction of brain parenchyma. To compensate for the resulting changes in intracranial pressure, cerebrospinal fluid volume increases (33). aSAH also induces hypokalemia via catecholamine-mediated β2-adrenergic stimulation of Na⁺/K⁺-ATPase, resulting in intracellular potassium shift (34–36).

Sodium ions are the dominant component of plasma osmolarity. Their acute elevation often leads to radiographically apparent ventricular enlargement, which is typically associated with low or normal intracranial pressure and can be reversed upon correction of the sodium imbalance. This condition is easily misdiagnosed as shunt failure in clinical settings, warranting particular caution in shunt-dependent patients (37, 38). Mechanistically, blood hyperosmolarity drives water movement out of brain cells, reducing brain volume and tissue tension while increasing intracranial compliance. The resulting compensatory space is occupied by CSF, leading to osmotic ventricular enlargement—a process consistent with the rapid reduction of intracranial pressure observed with hypertonic saline administration (39, 40). At the microscopic level, the choroid plexus establishes transepithelial electrochemical and osmotic gradients through Na⁺/K⁺-ATPase, carbonic anhydrase, and the Na⁺-K⁺-2Cl⁻ cotransporter, facilitating water influx into the ventricles via channels such as AQP1 and generating nearly iso-osmotic CSF. Animal and molecular studies indicate that AQP1 deficiency reduces CSF production and intracranial pressure, while NKCC1-mediated ion-water cotransport significantly contributes to CSF formation (41–43). Therefore, hypernatremia-associated ventricular enlargement should be managed through timely correction of electrolyte and osmotic imbalances.

These findings were consistent with our study, suggesting that hypernatremia in aSAH may contribute to the development of chronic hydrocephalus. Therefore, electrolyte levels should be closely monitored during hospitalization, and any imbalances should be promptly corrected.

Lactic acid elevation in CSF can be caused by various central nervous system disorders, including acute intracranial infections, stroke, SAH, seizures, and mitochondrial diseases (44–46). In aSAH, obstruction of fourth ventricular CSF outflow, subarachnoid erythrocytosis, and hemolysis elevate CSF lactate levels. Consistent with our findings, previous studies have shown that increased CSF lactate levels in the subarachnoid space predict the development of hydrocephalus (45). In addition, lactic acid may be produced within brain tissue and diffuse into the CSF space, explaining the difference in lactate levels between blood and CSF. High levels of extracellular lactate are primarily related to hyperglycolysis in the brain rather than hypoxia (46). Some studies suggest that the increased lactate concentration in CSF is related to protein and white blood cell count, indicating that local inflammation and blood–brain barrier dysfunction may play a role in increased CSF lactate levels (47, 48). Other research has shown that preoperative CSF glucose levels <2.8 mmol/L and CSF lactate levels >2.8 mmol/L are risk factors for infection after neurosurgery (47, 48). Furthermore, intracranial infection can affect the function of glucose transporters, decreasing the transport of glucose from blood to CSF. Under normal circumstances, the lactate content in the CSF is low, as it is mainly a product of anaerobic glucose fermentation within the CSF. CSF lactate levels are not easily influenced by peripheral blood lactate levels, making it a reliable indicator of metabolic state in the brain. High glucose levels in CSF can lead to fluid transfer from brain parenchymal cells to the extracellular space, resulting in ventricular enlargement (21, 49). However, this study did not find a significant correlation between glucose levels and chronic hydrocephalus. Therefore, CSF lactate is a more reliable biomarker for predicting chronic hydrocephalus post-aSAH.

Studies by Hu et al. and Gluski et al. have shown that the total volume of intracerebral hemorrhage is closely related to acute hydrocephalus (57, 58). A study by Mijderwijk et al. demonstrated that a high Fisher grade at admission was associated with chronic hydrocephalus (15). This association occurred because bleeding triggered an inflammatory response that stimulated the arachnoid membrane, leading to substantial protein deposition and abnormal cerebrospinal fluid circulation (17). In addition, a high Fisher grade is often linked to a longer hospital stay (59). Our findings were consistent with these studies, showing that larger hemorrhage volumes could lead to slower hematoma absorption, increasing the risk of developing chronic hydrocephalus.

Severe SAH, characterized by extensive and irregular morphology, makes the traditional Fisher grade susceptible to inter-observer variability (60). It is difficult for clinicians to quantitatively assess the irregular shape and extent of the hemorrhage. In contrast, the 3D-Unet automatic segmentation model we developed provides precise quantification of the hemorrhage, providing clinicians with a clearer understanding of the patient’s condition rather than relying on a subjective grade assessment. This approach helps clinicians make better-informed decisions regarding the need for drainage surgery. Our 3D-Unet segmentation model provides objective quantification of hemorrhage volume, facilitating accurate risk stratification for drainage surgery.