Machine learning in developing a predictive model for chronic hydrocephalus following aneurysmal subarachnoid hemorrhage.

Objective: Using machine learning (ML) algorithms integrated with deep learning and radiomics technologies, we developed a nomogram model through an in-depth analysis and mining of clinical data and imaging features from patients with aneurysmal subarachnoid hemorrhage (aSAH). This model was aimed to predict the risk of developing chronic hydrocephalus in aSAH patients. Methods: This study enrolled 410 patients diagnosed with subarachnoid hemorrhage (SAH) in the Neurosurgery Department of the Affiliated People's Hospital of Jiangsu University between January 2020 and December 2023. Clinical and imaging characteristic data were collected from these patients. Using radiomic methods, we extracted features from the white matter surrounding the anterior horns of both lateral ventricles, ultimately selecting seven radiomic features to calculate the radiomics score. An automatic segmentation model based on the 3D-Unet architecture was specifically developed to measure hematoma volume. Initially, univariate analysis was conducted on all features, and the least absolute shrinkage and selection operator (LASSO) regression model was applied for feature selection using 10-fold cross-validation to optimize the penalty parameter. Key risk factors were identified, and various ML algorithms were used to construct and validate a predictive model, leading to the development of a clinical-radiological nomogram. To evaluate the model's discriminative ability, we performed receiver operating characteristic (ROC) curve analysis and calculated the area under the curve (AUC). Additionally, the consistency between model predictions and actual outcomes was assessed using calibration curves. Further evaluation included plotting precision-recall (P-R) curves, decision curve analysis (DCA), and clinical impact curves (CIC) to demonstrate the net benefit of the model at various thresholds in the training and test sets, validating its clinical utility. Results: A total of 180 patients were included, and a 3D-Unet automatic segmentation model was developed to accurately identify and quantify SAH volume. In the test set, the model achieved a Dice similarity coefficient (DSC) of 0.85 ± 0.04, an intersection over union (IoU) of 0.74 ± 0.06, a Hausdorff distance (HD) of 20.4 ± 12.3, and an average symmetric surface distance (ASSD) of 0.31 ± 0.23, demonstrating excellent performance in identifying SAHs. After screening features such as hematoma volume and radiomic score through univariate logistic regression (LR), 21 potential risk factors were identified. LASSO regression further refined these to nine key risk factors. Combining the results from both analyses, 6 independent predictive factors were determined: cerebrospinal fluid lactic acid level, sodium (Na), corpus callosum angle, interval to blood clearance, periventricular white matter changes, and hematoma volume. Among 8 ML models, the LR model showed the best performance, with AUC values of 0.884 [95% confidence interval (CI): 0.826-0.942] in the training cohort and 0.860 (95% CI: 0.758-0.962) in the test cohort. The calibration curve of the LR model showed a high agreement between predicted probabilities and observed outcomes. Additionally, DCA and CIC analyses demonstrated significant net benefits across different risk thresholds, confirming high consistency between predictions and actual outcomes. Conclusion: The developed 3D-Unet automatic segmentation model accurately identified hematomas and calculated their volume, addressing the challenge of quantitatively assessing SAH volume in clinical practice. Hematoma volume, a key risk factor, was integrated with clinical and radiological features from computed tomography (CT) scans using ML methods to construct a clinical-radiological nomogram. This nomogram effectively predicted the development of chronic hydrocephalus in patients with aSAH.