Cerebrospinal fluid-based clinical-radiomics model for predicting treatment prognosis of acute ischemic stroke

Introduction

Ischemic stroke (IS) is the main form of stroke, constituting approximately 70% of all strokes. It occurs due to acute arterial occlusion, leading to a sudden reduction in cerebral blood flow and oxygen supply (1). Early recanalization of occluded vessels is therefore the cornerstone of acute ischemic stroke (AIS) management (2). Thrombolytic and endovascular treatment have been the two most effective reperfusion therapies for AIS (3,4). However, the efficacy of thrombolytic treatment varies considerably among patients, with divergent clinical outcomes (5). Moreover, a substantial proportion of patients are at increased risk of hemorrhagic transformation and subsequent cerebral edema after thrombolysis, often leading to poor prognosis (6,7). Therefore, accurate early prognostic prediction for AIS patients undergoing thrombolysis is crucial for risk stratification and personalized treatment planning.

Previous studies have indicated the potential role of early cerebrospinal fluid (CSF) volume alterations in predicting AIS outcomes. Specifically, when IS occurs, neuroinflammation and blood-brain barrier disruption led to dynamic CSF changes, with its neuroinflammatory biomarkers and volume metrics showing correlations with infarct size, the National Institute of Health stroke scale (NIHSS) scores, and clinical outcomes (8-11). However, most existing research has focused on CSF volumetric analysis for predicting malignant ischemic infarction and cerebral edema (12-15), while largely overlooking the potential prognostic value of the sophisticated texture and density characteristics of CSF. Furthermore, many current studies include mixed cohorts of AIS patients receiving either thrombolytic or endovascular therapy, rather than specifically targeting the thrombolysis population. These limitations highlight the need for a more precise and noninvasive approach to quantify CSF changes for prognostic prediction in AIS patients undergoing thrombolytic treatment.

Recent advances in artificial intelligence have expanded into stroke care, offering valuable guidance for early stroke management (16). Radiomics, as an auxiliary diagnostic technique, enables high-throughput extraction of a large amount of reproducible image features from medical images, capturing subtle patterns beyond human visual perception. Several radiomics studies have focused on infarction lesions (17-21), affected brain regions (22), penumbra areas on magnetic resonance imaging (MRI) (23), or thrombus area on non-contrast computed tomography (NCCT) and computed tomography angiography (CTA) (24,25) to predict AIS outcomes. However, to our knowledge, the potential of CSF-based radiomics features for AIS prognosis prediction has not been adequately explored.

Given the established association between CSF dynamics and AIS pathogenesis, we hypothesize that radiomics analysis of the CSF regions may provide valuable prognostic information. Therefore, this study aimed to develop and validate a machine learning (ML) model that integrates clinical factors and CSF-based radiomics features to predict functional outcomes at discharge—as measured by the modified Rankin Scale (mRS)—in AIS patients receiving intravenous thrombolysis (IVT). An independent test cohort was included to evaluate model generalizability. This approach may facilitate early risk stratification and support personalized clinical management, ultimately improving functional outcomes. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2684/rc).

Methods

Study population

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was approved by the Ethics Committee of the Second Affiliated Hospital of Soochow University (No. JD-HG-2024-071), and the requirement of patient approval or written informed consent for reviewing medical records or images was waived by the Ethics Committee in view of the retrospective nature based on anonymized radiology reports. This study was retrospective and observational in nature, in accordance with relevant regulations and Ethics Committee guidance; clinical trial registration is not applicable to this study. In addition, there is no patient or public involvement during the design, conduct, reporting, interpretation or dissemination of the study. The data that support the findings of this study are available from the corresponding author upon reasonable request.

A total of 398 patients aged ≥18 years within 6 hours after onset with confirmed diagnosis of AIS based on cranial NCCT were consecutively and retrospectively enrolled from July 2018 to June 2023. All participants underwent IVT, with a subset receiving additional endovascular treatment (such as arterial thrombolysis or mechanical thrombectomy) within 6 hours after onset in the Second Affiliated Hospital of Soochow University. Patients admitted to hospital within 4.5–6 hours from onset should undergo CTA or computed tomography perfusion (CTP) before thrombolysis to confirm that the penumbra area is sufficient and meets the conditions for thrombolysis, according to the inclusion criteria of EXTEND trial with a definition of variable tissue (ratio 1.2 or 10 mL difference, <70 mL core) (26). All patients had complete preoperative computed tomography (CT) images and clinical data. This study excluded patients with cerebral infarction sequelae and mRS score >2 before this onset (n=13), those who underwent cranial CT at other hospitals or had contrast-enhanced CT without NCCT scans (n=7), and those with severe radiologist-confirmed metal artifacts, motion artifacts, or significant noise in the images (n=31). Additionally, another 11 patients were excluded due to blurred, incomplete, or missing CT sequence images. Moreover, patients with final diagnosis of transient ischemia attack, pregnancy, known malignancy, and occupying or hemorrhagic lesions on cranial CT were also ineligible (n=28). Finally, 308 patients were enrolled in this study, which were randomly divided into training and test cohort at a ratio of 4:1 (Figure 1).

Figure 1 Flowchart of exclusion and inclusion of patients in our study. AIS, acute ischemic stroke; CT, computed tomography; CTA, computed tomography angiography; CTP, computed tomography perfusion; IVT, intravenous thrombolysis; mRS, modified Rankin Scale.

Clinical information was systematically collected from medical records, including gender, age, admission time, recanalization therapies, onset-to-CT-examination time, onset-to-treatment time (OTT time), NIHSS grading at admission, stroke subtypes, hypertension, hyperglycemia, hyperlipidemia, atrial fibrillation, previous stroke and smoking history. Recanalization treatment was categorized as IVT alone (treatment1), IVT followed by arterial thrombolysis (treatment2), and IVT followed by mechanical thrombectomy (treatment3). NIHSS grading at admission classified the stroke severity as 1 (mild, NIHSS 0–4), 2 (moderate, NIHSS 5–15) and 3 (severe, NIHSS ≥16). Stroke subtypes were classified using ASCO criteria (A for atherosclerosis, S for small vessel disease, C for cardiac source and O for other cause). Moreover, The CSF volume was automatically calculated from pre-treatment cranial CT scans: first, the number of voxels in the CSF region was counted; second, the volume of an individual voxel was calculated by multiplying the voxel dimensions along the x, y and z axes; finally, the total CSF volume was obtained by multiplying the number of CSF voxels by the single voxel volume.

Functional outcomes in AIS patients were assessed at discharge using the mRS by a trained neurologist blinded to clinical outcomes and imaging data. The mRS scores were dichotomized into good (mRS ≤2, coded as 1) and adverse (mRS >2, coded as 0) outcomes, which served as the dependent variable in the prediction model. The at-discharge mRS was selected over the conventional 90-day mRS endpoint because it more directly captures early outcomes mediated by acute-phase pathophysiological processes and supports timely clinical decision-making for discharge planning and rehabilitation initiation.

Image acquisition and parameters

A preoperative NCCT examination was performed to obtain a whole head scan in each patient, ranging from the base to the apex of the skull. All CT scans were performed using any of four CT scanners with different parameters, including GE MEDICAL SYSTEMS BrightSpeed, Philips iCT 256, GE MEDICAL SYSTEMS Optima CT620 and UIH uCT 530. All scanners operated at a fixed tube voltage of 120 kV to minimize energy-dependent attenuation differences. Besides, the selected scanners demonstrated minimal slice thickness variation (0.5–0.625 mm), conforming to clinical standards for cranial CT examinations. All images were isotropically resampled to 1 mm³ voxels to ensure spatial resolution consistency. In addition, scanners with slice thickness >1 mm, variable tube voltage (80–140 kV), or non-standard reconstruction kernels were excluded to control potential sources of technical variability. Detailed specifications for the CT scans are provided in Table S1 in the supplemental material.

Study framework

The framework of our method is illustrated in Figure 2. Initially, the region of CSF was automatically segmented on each CT sequence to define the region of interest (ROI) for our study. Simultaneously, we collected 16 clinical characteristics. Subsequently, the radiomics features were extracted quantitatively from the CSF region. Following this, the minimum redundancy maximum relevance (mRMR) algorithm and the least absolute shrinkage and selection operator (LASSO) regression were utilized to select the top-ranking radiomics features, while statistical analysis combined with univariate and multivariate logistic regression (LR) analysis were performed to identify the most distinctive clinical features. The previously selected features were then respectively applied to construct the radiomics model and clinical model, followed by the integrating these features to develop a combined model. A detailed description of the proposed workflow is presented below, encompassing CSF segmentation, feature extraction, feature selection, model development and model analysis (Figure 2).

Figure 2 The study workflow of radiomics analysis. CSF, cerebrospinal fluid; DCA, decision curve analysis; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; k, the number of features to be selected; mRMR, minimum redundancy maximum relevance; NuSVC, Nu support vector classification; RF, random forest; ROC, receiver operating characteristic.

CSF segmentation

The CSF segmentation was automatically performed on head CT scans using a hybrid processing pipeline implemented in both MATLAB R2018b and Python 3.6.13. The procedural pipeline comprised the following steps: firstly, the intracranial area was isolated through skull stripping from the raw CT images using Image Processing Toolbox of MATLAB, and the resulting intracranial mask was binarized to intracranial area (marked as 1) and extracranial area (marked as 0). The remaining processing steps were implemented in Python using SimpleITK (version 2.1.1.2) and Numpy (version 1.19.5) libraries. An automatic threshold segmentation with intensity range of 0–15 Hounsfield units (HU) was applied to the original CT data to initially identify low-density regions indicative of CSF. This thresholded output was then intersected with the intracranial mask to obtain a coarse CSF segmentation confined to the intracranial vault. Furthermore, a hole-filling operation was applied to fill enclosed voids with intensity values below 20 HU within the segmented regions. This threshold was determined through clinician’s prior anatomical knowledge and validated by preliminary segmentation testing on a development subset to optimally distinguish CSF-filled voids from low-density brain tissue, resulting in the final refined CSF segmentation. Finally, two radiologists with respective diagnostic experience of 6 and 8 years reviewed these segmentation results independently without access to clinical information or mRS results.

Radiomics feature extraction

The CSF region of NCCT images was quantitatively analyzed using AIMS based on PyRadiomics (version 3.0.1) (27), an open-source python package for extracting radiomics features from medical imaging, which was referenced from the instructions of the image biomarker standardization initiative (IBSI) (28). A total of 1,874 quantitative radiomics features were extracted, encompassing seven categories: 19 first-order gray-level statistic features, 16 3D shape-based features, 24 gray-level co-occurrence matrix (GLCM) features, 16 gray-level run length matrix (GLRLM) features, 16 gray-level size zone matrix (GLSZM) features, 5 neighboring gray tone difference matrix (NGTDM) features, and 14 gray-level dependence matrix (GLDM) features (see Figure S1 for details). In addition, eight image filters, including Laplacian of Gaussian, wavelet, logarithm, square, square root, exponential, gradient, and 3D local binary pattern, were applied to each original image to derive transformational image features. Moreover, the wavelet transform generated eight combinations of high-pass (H) and low-pass (L) filters in three dimensions, including LLH, LHL, LHH, HLL, HLH, HHL, HHH, and LLL.

Radiomic and clinical feature selection

In the procedure of radiomics feature selection, a two-stage strategy combining mRMR and LASSO was employed to enhance feature robustness and model interpretability by leveraging their complementary advantages: mRMR preselects features by maximizing relevance and minimizing redundancy, while LASSO further refines the subset by introducing sparsity and improving generalizability. Firstly, all radiomics features were normalized to a certain range of 0 to 1 through Z-score scaling, with the scaling parameters being fit solely on the training cohort to prevent data leakage, before being applied to the test set. This process was implemented to eliminate value range discrepancies between data and mitigate the influence of outliers. Secondly, mRMR algorithm was applied to identify the top 50 features that were highly predictive of at-discharge mRS scores while maintaining low inter-feature redundancy. Thirdly, the top features selected by mRMR were further refined using LASSO regression, which applies an L1 penalty to shrink the coefficients of irrelevant features to zero. The regularization parameter λ, which controls the degree of sparsity, was optimized through five-fold cross-validation. This yielded a sparse subset of features with nonzero coefficients that constitute the final radiomics signature.

Concurrently, the clinical features with strong support were selected by employing statistical analysis combined with univariate and multivariate LR analysis. The clinical characteristics with statistical significance in the univariate LR were further included in the multivariate LR to identify the ultimate predictor variables for constructing the clinical model. Odds ratios (ORs) with 95% confidence intervals (CIs) were calculated for each of these factors. The feature selection procedures were performed on the training cohort and used for the test cohort.

Model development and analysis

The flow of dataset division and model training is shown in Figure S2. This study encompassed a total of 308 patients, with 155 exhibiting good mRS scores and 153 showing otherwise. All patients were randomly divided into a primary training cohort (n=246) and a test cohort (n=62).

The previously selected radiomics features and clinical features were individually utilized to construct radiomics and clinical models, followed by the concatenating these features to develop a combined model. All of the models were trained and validated under three mainstream ML classifiers, including Nu support vector classification (NuSVC), LR, and random forest (RF). Hyperparameter optimization for these classifiers was conducted via a grid search with an inner five-fold cross-validation approach on the training cohort. The random seed was fixed at 42 throughout the process to ensure reproducibility. The hyperparameter grids for each classifier were defined as follows: NuSVC {nu: [0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55, 0.6}; gamma: [2⁻¹⁰, 2⁻⁹, 2⁻⁸, 2⁻⁷, 2⁻⁶, 2⁻⁵, 2⁻⁴, 2⁻³, 2⁻², 2⁻¹, 2⁰, 2¹, 2², 2³, 2⁴], LR {penalty: ['l1', 'l2', 'elasticnet']; C: [0.001, 0.01, 0.1, 1, 10, 100]}, and RF {n_estimators: [50, 75, 100, 125, 150, 175, 200]; max_depth ={3, 5, 10, 15, 20, None]}. The optimal hyperparameter combination for each model was selected based on the highest mean cross-validation AUC. Using these optimal parameters, a final model was refit on the entire training cohort and then evaluated on the independent test cohort to assess its performance in predicting the at-discharge mRS classification.

The ML model development and decision curve analysis (DCA) were performed using scikit-learn (version 0.24.2), an open-source Python library (29). The predictive performances of the models would be evaluated with standard metrics, including the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, accuracy, sensitivity and specificity. Furthermore, DCA was employed to access the clinical utility of the combined clinical-radiomics model by quantifying the net benefits associated with different threshold probabilities. Additionally, calibration curves were plotted using “rms” package (version 6.8.1) of R software in order to evaluate precision of model calibration, providing a comprehensive assessment of the clinical performance in discharge mRS classification.

Statistical analysis

Statistical analyses were conducted with SPSS (version 23.0) and R project (version 4.4.1). To check the equality of patient clinical data between different cohorts, continuous variables were analyzed using either Student’s t-test or Mann-Whitney U test. Specifically, normally distributed data were subjected to a t-test and presented as mean ± standard deviation, while non-normally distributed data were evaluated using the Mann-Whitney U test and expressed as median (interquartile range). Categorical variables were compared using the Chi-squared test. The clinical characteristics with the P value less than 0.05 in univariate LR analysis, were input into the multivariate LR analysis. The clinical model was adjusted for age, arrival time, treatment types, NIHSS grading, ASCO classification, a history of hyperglycemia, hyperlipidemia and atrial fibrillation.

For the five-fold cross-validation AUCs, the 95% CIs were derived from the mean ± 1.96 × standard deviation of the AUCs across the five folds; for the hold-out test cohort AUCs, the 95% CI was computed using the DeLong method, a robust non-parametric statistical approach. Pairwise comparisons of AUCs were performed using DeLong’s test for correlated ROC curves. A two-sided P value less than 0.05 was considered statistically significant for all analyses unless otherwise noted. All analyses were conducted on a 64-bit Windows 10 operating system.

Results

Patient characteristics

A total of 308 AIS patients were ultimately recruited for our study, with 246 (124 good prognosis and 122 poor prognosis) in the training cohort and 62 (31 good prognosis and 31 poor prognosis) in the test cohort. Besides onset-to-CT-examination time (P=0.049), OTT time (P=0.035) and hyperlipidemia (P=0.023), there were no significant differences in other clinical characteristics between the training and test cohorts. Moreover, there was no highly significant difference observed in the aforementioned three features, with P values no less than 0.01 (Table S2). In the training cohort, patients with poor prognosis tended to be older, and have a higher proportion of non-atherosclerosis subtypes and subset receiving additional endovascular treatment, a higher admission NIHSS grading, a higher prevalence of hyperglycemia and atrial fibrillation, and a lower prevalence of hyperlipidemia (all P<0.05). While in test cohort, patients with poor prognosis tended to arrive at hospital between 3 pm and 7 am on the following day, and have a higher proportion of subset receiving additional endovascular treatment, a higher admission NIHSS grading and prevalence of hyperglycemia (all P<0.05) (Table 1).

Univariate and multivariate LR analysis identified NIHSS grading, ASCO classification and hyperglycemia as independent predictors of functional outcome at discharge in AIS patients undergoing IVT. These variables were subsequently incorporated into the clinical prediction models. The analysis demonstrated that a higher NIHSS grading (OR, 0.184; 95% CI: 0.105–0.321), the presence of hyperglycemia (OR, 0.494; 95% CI: 0.252–0.970), and non-atherosclerosis subtypes (OR, 0.556; 95% CI: 0.328–0.941) were significantly associated with reduced odds of a good functional outcome (Table 2).

Radiomics model construction

A total of 1,874 radiomics features were extracted from the CSF region of CT scan images. Among these features, 50 features were selected with mRMR, among them 21 features were further selected through LASSO regression with the Lamda value of 0.013 (Figure 3). Table S3 lists the LASSO coefficients, radiomics implication and corresponding formulas of 21 radiomics features, with positive LASSO coefficients indicating favorable prognosis and negative ones adverse. Furthermore, coefficient’s absolute value is positively correlated with the feature importance. Among three ML radiomics models (NuSVC, LR, RF) evaluated through five-fold cross-validation and test cohort, the radiomics models under NuSVC performed the best, achieving an AUC of 0.786, an accuracy of 0.710, a sensitivity of 0.729, a specificity of 0.690 and a precision of 0.705 through five-fold cross-validation. In the test cohort, the NuSVC model achieved an AUC of 0.820, an accuracy of 0.774, a sensitivity of 0.871, a specificity of 0.677, and a precision of 0.730 (refer to Figure 4 and Table 3 for details), making it the optimal ML radiomics model for our analysis.

Figure 3 LASSO coefficient profile plot with different log (λ), with the vertical dashed lines representing 21 radiomics features with non-zero coefficients selected under the λ value of 0.013 (A), and names and coefficients of the selected radiomics features (B). GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run length matrix; GLSZM, gray-level size zone matrix; H, high-pass; L, low-pass; LASSO, least absolute shrinkage and selection operator; NGTDM, neighboring gray tone difference matrix; ROI, region of interest.

Figure 4 The receiver operating characteristic curves of three machine learning classifiers through the five-fold cross-validation (A) and test cohort (B) results. AUC, area under the curve; LR, logistic regression; NuSVC, Nu support vector classification; RF, random forest.

Combined model construction and validation

The combined clinical-radiomics model was developed by integrating three clinical predictors with 21 radiomics features. To enable a uniform and objective comparison of diagnostic efficacy, the radiomics model, clinical model, and combined model were each constructed using three ML classifiers. The ROC curves for all models under the three classifiers, evaluated through five-fold cross-validation and on the test cohort, are shown in Figure 5. The result indicated that the combined model outperformed both the individual clinical and radiomics signature models alone (Table 3). Among all models, the combined model under NuSVC classifier achieved the best performance, with an AUC of 0.870 (95% CI: 0.850–0.889) in cross-validation and 0.893 (95% CI: 0.817–0.968) in hold-out test cohort. Pairwise model comparisons across all classifiers using DeLong tests are summarized in Table S4. Specifically, for the NuSVC-based models, the radiomics and clinical models showed comparable performance (P=0.906). Although the combined model exhibited a statistically significant improvement over the radiomics model (P=0.036), its enhancement compared to the clinical model did not reach statistical significance (P=0.105) in the test cohort.

Figure 5 The receiver operating characteristic curves of the radiomics model, clinical model, and combined model with three machine learning classifiers in the five-fold cross-validation (A-C) and test cohort (D-F). AUC, area under the curve; LR, logistic regression; NuSVC, Nu support vector classification; RF, random forest.

The DCA for the radiomics model, clinical model and the combined model under NuSVC classifier is presented in Figure 6A. The DCA evaluates the clinical utility of the models by quantifying the net benefit across various threshold probabilities (x-axis), which represents the lowest risk probability that a patient or a doctor is willing to accept intervention in clinical decision-making. The y-axis represents the net benefit, balancing true positives and false positives to determine whether using the model improves decision-making compared to treating all or none as positive. The combined model (red line) provides a greater net benefit over a wide range of thresholds compared to both the radiomics model (blue line) and clinical model (green line), as well as when all patients are treated (gray line) or none of them are treated as poor prognosis (gray dash line). These findings highlight the favorable clinical utility of the combined model in classifying functional outcomes and guiding personalized treatment decisions for AIS patients.

Figure 6 Decision curve analysis (A) and calibration curves (B) of three NuSVC models in the test cohort. NuSVC, Nu support vector classification.

The calibration curves of the clinical model, radiomics model and combined model with NuSVC classifier are depicted in Figure 6B. In the test cohort, the Hosmer-Lemeshow test for the NuSVC combined model resulted in a Chi-squared value of 6.51 and a nonsignificant P value of 0.59 (>0.05), indicating that there is no statistically significant difference between the predicted and observed values, thus demonstrating favorable calibration power of the prediction model. Furthermore, compared with the radiomics model (blue line), radiomics-clinical combined model (red line) exhibited superior alignment with the ideal line and enhanced calibration precision.

Discussion

In this study, a moderate sample size of 308 AIS patients within 6 hours from onset was included in the final analysis, with 246 and 62 patients allocated to the training and test sets respectively. This study extracted radiomics features from the CSF region in CT scan images, and constructed a combined model which integrated clinical and radiomics characteristics to predict the functional outcomes of AIS patients undergoing thrombolysis treatment. Encouragingly, through integrating three clinical features and 21 shape features, the combined model exhibited improved prognostic performance and enhanced clinical utility compared with either the clinical or radiomics model alone. Specifically, the combined model under the optimal classifier NuSVC achieved an AUC of 0.893, an accuracy of 0.774, a sensitivity of 0.806, a specificity of 0.742 and a precision of 0.758 for at-discharge functional outcome prediction, indicating its favorable predictive ability. In addition, the combined model exhibited greater net benefit and good calibration according to the calibration curve analysis and DCA. Thus, this study could offer a viable and trustworthy non-invasive method for predicting functional prognosis in AIS patients and facilitating early risk stratification and timely intervention for better functional outcomes.

The successful recanalization rate for thrombolysis falls below 40% (30), leading to a significant proportion of patients experiencing unfavorable functional outcomes. Due to the complex mechanism of functional prognosis after thrombolysis combined with or without endovascular treatment, the research on prediction of AIS prognosis is still controversial. Because of the sensitivity and effectiveness of MRI imaging detection for AIS, most previous researches on radiomics analysis for predicting clinical outcome of AIS patients have focused on infarction lesions (17-21) or penumbra area (22,23) based on pre-treatment MRI images. However, the high cost, long examination duration and low popularity of emergency MRI made the pre-treatment MRI examination within the limited treatment time window significantly difficult. Considering both the significant contribution of CSF region to AIS progression and the popularity of CT imaging, our study proposed a CT-based radiomics-clinical model by extracting imaging information in the CSF region for prognosis prediction of hyper-acute AIS patients. Currently, due to the patient heterogeneity and the uncertainty of the treatment benefit, it is still a technical bottleneck for clinicians to predict the clinical outcome following vascular recanalization treatments of AIS patients. To the best of our knowledge, our study is the first to introduce radiomics technology to evaluate the clinical and radiological associations in AIS patients based on CSF identification.

Previous studies have primarily emphasized the contribution of CSF volume changes in AIS prognosis. However, post-ischemic alterations in neuroinflammatory markers in CSF and other body fluids (including matrix metalloproteinase 9, C-reactive protein, interleukin 1β, interleukin 6 and interleukin 10) would either promote or inhibit neuroinflammation within the ischemic brain region, alter CSF properties and ultimately affect radiomics features (31,32). Moreover, elevated concentration of proteins and other bioactive molecules in the CSF after AIS were reported, which might also lead to the density and texture alteration of CSF (33). This study identified 21 CSF-based radiomics features highly correlated with AIS prognosis, encompassing four first-order histogram features, two shape features, six GLCM features, three GLSZM features, three GLDM features, two NGTDM features, and one GLRLM features. Among them, the Mean (the average gray level intensity of CSF) is positively correlated with favorable prognosis, indicating that increased density of proteins, growth factors, and other bioactive molecules within the CSF could be pivotal in promoting neural repair mechanisms within lesioned brain regions. This increased CSF density likely reflects elevated neurotrophic factors, including brain-derived neurotrophic factor (BDNF), glial cell line-derived neurotrophic factor (GDNF), and nerve growth factor (NGF), which enhance neuronal survival, synaptic plasticity, and functional recovery in damaged brain regions, ultimately contributing to a favorable prognosis (34-36). A higher inverse variance (the dispersion of data distribution) indicates more concentrated gray level data and homogeneous texture patterns, this homogeneity may reflect a transition from acute inflammatory phases—marked by blood-brain barrier disruption and biochemical heterogeneity—to a stabilization phase characterized by uniform protein deposition and attenuated cytokine activity, resulting in a favorable clinical outcome (37). A higher SizeZoneNonUniformityNormalized (the variability of size zone volumes throughout the image) suggests more heterogeneity among zone size volumes and fewer large connected domains, which leads to finer texture and a favorable clinical outcome. A higher busyness (the frequency of intensity changes between pixels and their neighbors in an image) represents more complex and uneven textures, leading to a worse clinical outcome. A higher sphericity (the roundness of the shape of the ROI region) indicated less flatness and a smaller surface area for a given volume, this spherical morphology possibly stems from post-ischemic ventricular remodeling, where peri-lesional tissue atrophy (e.g., white matter degeneration) triggers compensatory CSF space expansion, disrupting normal CSF flow patterns and reducing functional reserve capacity. Concomitant structural alterations in adjacent cisterns and gyri further exacerbate this loss of CSF region flatness, linking to an adverse clinical outcome (38). To summarize, favorable AIS prognosis is associated with higher CSF density, fewer large plaques, finer and more homogeneous texture patterns, and improved flatness in CSF region.

In this study, the combined model under NuSVC demonstrated excellent performance in predicting clinical outcomes with an AUC of 0.893 in the test cohort, which distinctly outperformed either the clinical model (AUC of 0.828) or radiomics model (AUC of 0.820) individually. This result highlighted the significant contribution of quantitative CSF radiomics features to functional outcome prediction in AIS patients by enhancing the performance of the clinical model. In addition, the incorporation of both clinical signatures and imaging features through ML methods in a combined model could significantly enhance the predictive performance. Compared with recent studies, our CSF-based model exhibited superior performance in predicting functional outcomes for AIS patients compared to Zhou’s study based on AIS lesion (AUC: 0.890) (17), Wang’s study based on infarction lesions (AUC: 0.73) (18), Quan’s study based on AIS lesions (AUC: 0.864) (19), Tang’s study based on penumbra (AUC: 0.886) (23), Ramos’s study based on 70 brain regions (AUC: 0.8) (39), and Jabal’s study based on brain regions of CT image (AUC: 0.84) (40). These results underscored the powerful predictive potential of CSF-based radiomics features for treatment prognosis prediction in AIS patients.

There were several limitations in this study. First, despite the model’s robust performance among multiple devices and parameters, the single-center retrospective design and use of a single imaging modality may limit generalizability. The sample size may also be insufficient to capture the full heterogeneity of prognosis in AIS populations. Second, the lack of long-term clinical follow-up might affect prognostic reliability. Third, the exclusion of 31 cases due to severe motion or metal artifacts may introduce a “clean data selection bias” and affect model generalizability in real-world clinical settings. Finally, the limited sample size prevented the application of deep learning approaches in this study. Based on the above limitations, the following future directions are proposed: (I) multicenter studies with larger cohorts are warranted to verify prognostic robustness and enhance model generalizability in predicting functional outcomes in AIS patients; (II) standardized long-term outcome assessments could be incorporated, with Kaplan-Meier survival analysis to compare survival differences between different treatment strategies; (III) synthetic data augmentation as well as deep learning-based motion artifact correction and metal artifact reduction could be performed for improving model robustness to artifacts; (IV) the synergistic integration of deep learning techniques and radiomics analysis could be prioritized to improve diagnostic precision in AIS patients.

Conclusions

In conclusion, our study proposed and verified a CSF-based combined model incorporating clinical features with radiomics features to predict the at-discharge prognosis of AIS patients. The result demonstrated that the combined clinical-radiomics model using NuSVC classifier exhibited superior diagnostic performance in clinical prognosis prediction. This innovative approach holds promise for risk stratification and personalized intervention guidance in AIS patients.

Acknowledgments

The authors wish to acknowledge the contributing research center delivering data to this research, all of the study participants and their relatives and the clinical staff for their support and contribution to this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2684/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2684/dss

Funding: This work was supported by the National Natural Science Foundation of China (62441114); Suzhou Key Laboratory Project (No. SZS2024007); Suzhou Science & Technology Projects (Nos. SKY2022151, SKY2023070 and SKY2023179); and Key Talent Project of Nuclear Technology Medical Application of the Second Affiliated Hospital of Soochow University (No. XKTJ-HRC2021006). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the article.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2684/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was approved by the Ethics Committee of the Second Affiliated Hospital of Soochow University (No. JD-HG-2024-071), and the requirement of patient approval or written informed consent for reviewing medical records or images was waived by the Ethics Committee in view of the retrospective nature based on anonymized radiology reports.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Darwish EAF, Abdelhameed-El-Nouby M, Geneidy E. Mapping the ischemic penumbra and predicting stroke progression in acute ischemic stroke: the overlooked role of susceptibility weighted imaging. Insights Imaging 2020;11:6. [Crossref] [PubMed]
2. Phipps MS, Cronin CA. Management of acute ischemic stroke. BMJ 2020;368:l6983. [Crossref] [PubMed]
3. Berge E, Whiteley W, Audebert H, De Marchis GM, Fonseca AC, Padiglioni C, de la Ossa NP, Strbian D, Tsivgoulis G, Turc G. European Stroke Organisation (ESO) guidelines on intravenous thrombolysis for acute ischaemic stroke. Eur Stroke J 2021;6:I-LXII. [Crossref] [PubMed]
4. Liu L, Chen W, Zhou H, Duan W, Li S, Huo X, Xu W, Huang L, Zheng H, Liu J, Liu H, Wei Y, Xu J, Wang YChinese Stroke Association Stroke Council Guideline Writing Committee. Chinese Stroke Association guidelines for clinical management of cerebrovascular disorders: executive summary and 2019 update of clinical management of ischaemic cerebrovascular diseases. Stroke Vasc Neurol 2020;5:159-76. [Crossref] [PubMed]
5. Grotta JC. Intravenous Thrombolysis for Acute Ischemic Stroke. Continuum (Minneap Minn) 2023;29:425-42. [Crossref] [PubMed]
6. He J, Fu F, Zhang W, Zhan Z, Cheng Z. Prognostic significance of the clinical and radiological haemorrhagic transformation subtypes in acute ischaemic stroke: A systematic review and meta-analysis. Eur J Neurol 2022;29:3449-59. [Crossref] [PubMed]
7. Qiu L, Fu F, Zhang W, He J, Zhan Z, Cheng Z. Prevalence, risk factors, and clinical outcomes of remote intracerebral hemorrhage after intravenous thrombolysis in acute ischemic stroke: a systematic review and meta-analysis. J Neurol 2023;270:651-61. [Crossref] [PubMed]
8. Jayaraj RL, Azimullah S, Beiram R, Jalal FY, Rosenberg GA. Neuroinflammation: friend and foe for ischemic stroke. J Neuroinflammation 2019;16:142. [Crossref] [PubMed]
9. Maida CD, Norrito RL, Daidone M, Tuttolomondo A, Pinto A. Neuroinflammatory Mechanisms in Ischemic Stroke: Focus on Cardioembolic Stroke, Background, and Therapeutic Approaches. Int J Mol Sci 2020;21:6454. [Crossref] [PubMed]
10. Smith CJ, Hulme S, Vail A, Heal C, Parry-Jones AR, Scarth S, Hopkins K, Hoadley M, Allan SM, Rothwell NJ, Hopkins SJ, Tyrrell PJ. SCIL-STROKE (Subcutaneous Interleukin-1 Receptor Antagonist in Ischemic Stroke): A Randomized Controlled Phase 2 Trial. Stroke 2018;49:1210-6. [Crossref] [PubMed]
11. Mestre H, Du T, Sweeney AM, Liu G, Samson AJ, Peng W, et al. Cerebrospinal fluid influx drives acute ischemic tissue swelling. Science 2020;367:eaax7171. [Crossref] [PubMed]
12. Kauw F, Bennink E, de Jong HWAM, Kappelle LJ, Horsch AD, Velthuis BK, Dankbaar JW. Intracranial Cerebrospinal Fluid Volume as a Predictor of Malignant Middle Cerebral Artery Infarction. Stroke 2019;50:1437-43. [Crossref] [PubMed]
13. Kauw F, Bernsen MLE, Dankbaar JW, de Jong HW, Kappelle LJ, Velthuis BK, van der Worp HB, van der Lugt A, Roos YB, Yo LS, van Walderveen MA, Hofmeijer J, Bennink E. Cerebrospinal fluid volume improves prediction of malignant edema after endovascular treatment of stroke. Int J Stroke 2023;18:187-92. [Crossref] [PubMed]
14. Foroushani HM, Hamzehloo A, Kumar A, Chen Y, Heitsch L, Slowik A, Strbian D, Lee JM, Marcus DS, Dhar R. Quantitative Serial CT Imaging-Derived Features Improve Prediction of Malignant Cerebral Edema after Ischemic Stroke. Neurocrit Care 2020;33:785-92. [Crossref] [PubMed]
15. Dhar R, Chen Y, Hamzehloo A, Kumar A, Heitsch L, He J, Chen L, Slowik A, Strbian D, Lee JM. Reduction in Cerebrospinal Fluid Volume as an Early Quantitative Biomarker of Cerebral Edema After Ischemic Stroke. Stroke 2020;51:462-7. [Crossref] [PubMed]
16. Fan Y, Song Z, Zhang M. Emerging frontiers of artificial intelligence and machine learning in ischemic stroke: a comprehensive investigation of state-of-the-art methodologies, clinical applications, and unraveling challenges. EPMA J 2023;14:645-61. [Crossref] [PubMed]
17. Zhou Y, Wu D, Yan S, Xie Y, Zhang S, Lv W, Qin Y, Liu Y, Liu C, Lu J, Li J, Zhu H, Liu WV, Liu H, Zhang G, Zhu W. Feasibility of a Clinical-Radiomics Model to Predict the Outcomes of Acute Ischemic Stroke. Korean J Radiol 2022;23:811-20. [Crossref] [PubMed]
18. Wang H, Sun Y, Ge Y, Wu PY, Lin J, Zhao J, Song B. A Clinical-Radiomics Nomogram for Functional Outcome Predictions in Ischemic Stroke. Neurol Ther 2021;10:819-32. [Crossref] [PubMed]
19. Quan G, Ban R, Ren JL, Liu Y, Wang W, Dai S, Yuan T. FLAIR and ADC Image-Based Radiomics Features as Predictive Biomarkers of Unfavorable Outcome in Patients With Acute Ischemic Stroke. Front Neurosci 2021;15:730879. [Crossref] [PubMed]
20. Li Y, Liu Y, Hong Z, Wang Y, Lu X. Combining machine learning with radiomics features in predicting outcomes after mechanical thrombectomy in patients with acute ischemic stroke. Comput Methods Programs Biomed 2022;225:107093. [Crossref] [PubMed]
21. Yu H, Wang Z, Sun Y, Bo W, Duan K, Song C, Hu Y, Zhou J, Mu Z, Wu N. Prognosis of ischemic stroke predicted by machine learning based on multi-modal MRI radiomics. Front Psychiatry 2022;13:1105496. [Crossref] [PubMed]
22. Meng Y, Wang H, Wu C, Liu X, Qu L, Shi Y. Prediction Model of Hemorrhage Transformation in Patient with Acute Ischemic Stroke Based on Multiparametric MRI Radiomics and Machine Learning. Brain Sci 2022;12:858. [Crossref] [PubMed]
23. Tang TY, Jiao Y, Cui Y, Zhao DL, Zhang Y, Wang Z, Meng XP, Yin XD, Yang YJ, Teng GJ, Ju SH. Penumbra-based radiomics signature as prognostic biomarkers for thrombolysis of acute ischemic stroke patients: a multicenter cohort study. J Neurol 2020;267:1454-63. [Crossref] [PubMed]
24. Hofmeister J, Bernava G, Rosi A, Vargas MI, Carrera E, Montet X, Burgermeister S, Poletti PA, Platon A, Lovblad KO, Machi P. Clot-Based Radiomics Predict a Mechanical Thrombectomy Strategy for Successful Recanalization in Acute Ischemic Stroke. Stroke 2020;51:2488-94. [Crossref] [PubMed]
25. van Voorst H, Bruggeman AAE, Yang W, Andriessen J, Welberg E, Dutra BG, et al. Thrombus radiomics in patients with anterior circulation acute ischemic stroke undergoing endovascular treatment. J Neurointerv Surg 2023;15:e79-85. [Crossref] [PubMed]
26. Ma H, Campbell BCV, Parsons MW, Churilov L, Levi CR, Hsu C, et al. Thrombolysis Guided by Perfusion Imaging up to 9 Hours after Onset of Stroke. N Engl J Med 2019;380:1795-803. [Crossref] [PubMed]
27. Zhou Z, Qian X, Hu J, Chen G, Zhang C, Zhu J, Dai Y. An artificial intelligence-assisted diagnosis modeling software (AIMS) platform based on medical images and machine learning: a development and validation study. Quant Imaging Med Surg 2023;13:7504-22. [Crossref] [PubMed]
28. Zwanenburg A, Vallières M, Abdalah MA, Aerts HJWL, Andrearczyk V, Apte A, et al. The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology 2020;295:328-38. [Crossref] [PubMed]
29. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay É. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 2011;12:2825-30.
30. Shafie M, Yu W. Recanalization Therapy for Acute Ischemic Stroke with Large Vessel Occlusion: Where We Are and What Comes Next? Transl Stroke Res 2021;12:369-81. [Crossref] [PubMed]
31. Gao Y, Fang C, Wang J, Ye Y, Li Y, Xu Q, Kang X, Gu L. Neuroinflammatory Biomarkers in the Brain, Cerebrospinal Fluid, and Blood After Ischemic Stroke. Mol Neurobiol 2023;60:5117-36. [Crossref] [PubMed]
32. Zhang X, Bi X. Post-Stroke Cognitive Impairment: A Review Focusing on Molecular Biomarkers. J Mol Neurosci 2020;70:1244-54. [Crossref] [PubMed]
33. Rundblad LIS, Iversen HK, West AS. Pleocytosis in cerebrospinal fluid attributed to ischemic stroke: A review of the literature. J Neurol Sci 2023;449:120664. [Crossref] [PubMed]
34. Karantali E, Kazis D, Papavasileiou V, Prevezianou A, Chatzikonstantinou S, Petridis F, McKenna J, Luca AC, Trus C, Ciobica A, Mavroudis I. Serum BDNF Levels in Acute Stroke: A Systematic Review and Meta-Analysis. Medicina (Kaunas) 2021;57:297. [Crossref] [PubMed]
35. Zhang Z, Zhang N, Ding S. Reactive Astrocytes Release GDNF to Promote Brain Recovery and Neuronal Survival Following Ischemic Stroke. Neurochem Res 2025;50:117. [Crossref] [PubMed]
36. Norata D, Capone F, Motolese F, Marano M, Rossi M, Calandrelli R, Sacchetti M, Mantelli F, Di Lazzaro V, Pilato F. 1953-2023. Seventy Years of the Nerve Growth Factor: A Potential Novel Treatment in Neurological Diseases? Aging Dis 2024;16:2293-314. [Crossref] [PubMed]
37. Rodríguez-Yáñez M, Castillo J. Role of inflammatory markers in brain ischemia. Curr Opin Neurol 2008;21:353-7. [Crossref] [PubMed]
38. Currà A, Pierelli F, Gasbarrone R, Mannarelli D, Nofroni I, Matone V, Marinelli L, Trompetto C, Fattapposta F, Missori P. The Ventricular System Enlarges Abnormally in the Seventies, Earlier in Men, and First in the Frontal Horn: A Study Based on More Than 3,000 Scans. Front Aging Neurosci 2019;11:294. [Crossref] [PubMed]
39. Ramos LA, van Os H, Hilbert A, Olabarriaga SD, van der Lugt A, Roos YBWEM, van Zwam WH, van Walderveen MAA, Ernst M, Zwinderman AH, Strijkers GJ, Majoie CBLM, Wermer MJH, Marquering HA. Combination of Radiological and Clinical Baseline Data for Outcome Prediction of Patients With an Acute Ischemic Stroke. Front Neurol 2022;13:809343. [Crossref] [PubMed]
40. Jabal MS, Joly O, Kallmes D, Harston G, Rabinstein A, Huynh T, Brinjikji W. Interpretable Machine Learning Modeling for Ischemic Stroke Outcome Prediction. Front Neurol 2022;13:884693. [Crossref] [PubMed]