Abstract
Backgrounds
Futile recanalization (FR) poses a significant challenge in endovascular treatment and there is a lack of reliable predictive models for assessing treatment outcomes in stroke. The aim of this study is to develop a robust CT radiomics-clinical ensemble model that predicts FR in patients with acute ischemic stroke (AIS) following endovascular treatment (EVT) utilizing machine learning techniques.
Methods
This study enrolled 101 patients diagnosed with AIS who underwent successful EVT. A total of 946 radiomics features were, respectively, extracted from non-contrast CT (NCCT), contrast-enhanced CT (CECT), and various CT perfusion maps (CBF, CBV, MTT, and TTP) using PyRadiomics prior to the endovascular intervention. Demographic characteristics, along with baseline clinical, laboratory, and angiographic variables, were incorporated as clinical features in the model analysis. Feature engineering was performed using SelectKBest. Five traditional machine learning algorithms were employed for modeling. The dataset was randomly split into a training cohort (n = 71, 70%) and an internal validation cohort (n = 30, 30%). Receiver operating characteristic (ROC) curves were utilized to evaluate the performance of each model.
Results
Among the 101 patients, FR occurred in 66 individuals (65%), as determined by the modified Rankin Scale (mRS) at 90 days. The ensemble model integrating clinical data, NCCT, and CBV achieved the highest performance, with an area under the curve (AUC) of 0.918 using the CatBoost algorithm.
Conclusion
The multimodal CT radiomics-clinical ensemble machine learning model demonstrated excellent predictive capability for identifying FR in AIS patients with large vessel occlusion prior to EVT.
Keywords: CT radiomics, futile recanalization, machine learning, predictive model, stroke
1. Introduction
Endovascular therapy (EVT) has been endorsed by international guidelines for treating acute ischemic stroke (AIS) patients with large vessel occlusion (LVO) (Lin et al., 2019; Powers et al., 2019). The global implementation of these guidelines has led to improved clinical outcomes for stroke patients in recent years. Despite achieving successful recanalization (modified Thrombolysis in Cerebral Infarction [mTICI] 2b-3) in approximately 90% of cases with EVT, nearly half of the patients still experience an unfavorable functional outcome at 90 days (modified Rankin Scale [mRS] 3–6) (Goyal et al., 2016; Happi Ngankou et al., 2021; Jia et al., 2021). This situation is often referred to as futile recanalization (FR) or clinically ineffective reperfusion (CIR) (Hussein et al., 2010; Nie et al., 2023). FR presents significant challenges and confusion in the clinical management of AIS.
To prevent FR and identify risk factors prior to endovascular treatment, numerous potential predictors and predictive models have been proposed, showing promising potential (Brugnara et al., 2020; D’Anna et al., 2024; Da Ros et al., 2024; Su et al., 2021). However, many previous studies have relied on limited imaging or clinical data and traditional statistical methods, which hinder their ability to integrate multimodal parameters to explore the complex causes and mechanisms of FR and accurately predict outcomes. Therefore, it is essential to develop comprehensive methods for conducting a thorough analysis of various factors using diverse data modalities.
Radiomics is an innovative analytical technique that extracts extensive information regarding the characteristics of biological tissues from large imaging datasets, providing quantitative, high-dimensional, and mineable features for further analysis (Varghese et al., 2019). It has demonstrated exceptional performance in areas such as disease diagnosis and prognosis prediction, and shows considerable promise in stroke management (Chen et al., 2021; Luo et al., 2023). Machine learning, a core component of artificial intelligence, encompasses a set of computational methods that proficiently process complex patterns and nonlinear relationships within substantial datasets—capabilities that traditional statistical models struggle to achieve (Choy et al., 2018). Recent studies suggest that machine learning methods hold potential for predicting outcomes in AIS patients undergoing EVT; however, there is a lack of research applying radiomics and machine learning techniques specifically to predict FR (D’Anna et al., 2024; Da Ros et al., 2024).
This study aims to establish a preoperative ensemble model that effectively predicts FR using machine learning methods, incorporating multimodal radiomics features along with clinical, laboratory, and angiographic data. This approach seeks to enhance personalized management of stroke patients and ultimately identify those who would benefit most from endovascular treatment in clinical practice.
2. Materials and methods
This retrospective study was approved by the institutional ethics committee. The analysis included a consecutive cohort of 144 patients diagnosed with acute ischemic stroke (AIS) due to large vessel occlusion (LVO) who underwent endovascular treatment (EVT) between March 2018 and December 2023. Inclusion criteria comprised clinically confirmed cases of AIS with LVO treated at our hospital. Multimodal CT imaging, including non-contrast CT (NCCT), contrast-enhanced CT (CECT), and multi-parametric CT perfusion (CTP), was performed prior to the endovascular intervention. Comprehensive clinical, laboratory, and angiographic data were collected. Exclusion criteria involved cases with motion artifacts and those where the mTICI score was below 2b. “Incomplete clinical data” referred to missing key variables required for modeling or outcome assessment, including unavailable baseline clinical/laboratory records, incomplete angiographic timing variables, or missing 90-day mRS data. “Poor image quality” referred to severe motion artifact, incomplete imaging coverage, or unsatisfactory inter-modality alignment on visual quality control that precluded reliable radiomics extraction. Ultimately, 101 patients were included in the subsequent research analysis after excluding 15 patients with unsuccessful recanalization and mTICI < 2b, 20 patients with incomplete clinical data, and 8 patients with poor image quality. The patient selection process and overall study pipeline are illustrated in Figures 1, 2.
Figure 1.
Patient selection process and study workflow. AIS, acute ischemic stroke; LVO, large vessel occlusion; EVT, endovascular therapy; mTICI, modified thrombolysis in cerebral infarction; NCCT, non-contrast brain CT; CECT, contrast-enhanced CT; CTP, CT perfusion; CBF, cerebral blood flow; CBV, cerebral blood volume; MTT, mean transit time; TTP, time-to-peak; SVM, Support vector machine.
Figure 2.
Overall study pipeline.
2.1. Imaging protocols
All CT images were acquired on a 320-slice CT scanner (Aquilion ONE TSX-301A, Canon, Tokyo, Japan) with a standardized protocol. NCCT was performed first (matrix 512 × 512; typical 120 kVp; automatic mA modulation; slice thickness 1 mm). When clinically warranted, CECT was performed to evaluate vascular anatomy. Dynamic CTP was executed using a bolus-tracking protocol to create quantitative parametric maps: cerebral blood flow (CBF), cerebral blood volume (CBV), mean transit time (MTT), and time-to-peak (TTP). Perfusion maps were generated on the vendor workstation through the deconvolution of tissue time–attenuation curves. For the present radiomics analysis, only patients with complete multimodal datasets available for feature extraction were included in the final cohort. DICOM images from all six modalities (NCCT, CECT, CBF, CBV, MTT, and TTP) were exported for further analysis, adhering to the ALARA principles for dose management.
2.2. Image preprocessing and registration
Image preprocessing was conducted to ensure spatial and intensity consistency across all modalities before radiomics extraction. First, each modality was linearly registered to the subject’s NCCT in native space using FSL FLIRT (6–12 DOF affine with trilinear interpolation) to achieve sub-voxel alignment (Jenkinson et al., 2002). For CTP images, a time–space refinement process was applied before generation of parametric maps to reduce minor motion between dynamic frames. All co-registered images then underwent visual quality control by two imaging analysts to confirm acceptable alignment between lesion-containing regions across modalities; cases with unsatisfactory registration were excluded from further analysis. Prior to radiomics extraction, intensity clipping was applied to reduce the influence of outlier voxels, and z-score normalization was performed within the brain mask to improve cross-modality intensity comparability (Li et al., 2022; Pisani et al., 2025). Subsequently, during feature extraction, images were resampled to isotropic voxels and discretized according to the PyRadiomics/IBSI workflow described below.
2.3. Lesion segmentation
Two board-certified neuroradiologists, each with over 10 years of experience, independently delineated ischemic lesions using ITK-SNAP, referencing all modalities concurrently, including NCCT and perfusion maps. A consensus was reached for the final lesion mask for each case.
2.4. Radiomics feature extraction
Radiomics features were extracted using PyRadiomics (v3.0), following the guidelines established by the Image Biomarker Standardisation Initiative (IBSI) to ensure reproducibility and comparability (van Griethuysen et al., 2017). Before extraction, the volumes were resampled to create isotropic voxels, and the intensity values were discretized with a fixed bin width of 5 HU; gray-level normalization was applied to create 100 bins for weighted images, ensuring consistency in texture computation. For each imaging modality, we calculated first-order statistics, shape features, and second-order texture metrics, which included the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size-zone matrix (GLSZM), and gray-level dependence matrix (GLDM) (Figure 3). Higher-order features were generated from filtered images using the Laplacian of Gaussian filter with σ = 5 mm, alongside 3D wavelet sub-bands. In total, across both the original and ten filtered images, we extracted 946 features per modality (comprising 18 first-order, 22 GLCM, 16 GLRLM, 16 GLSZM, and 14 GLDM features per image type), resulting in a thorough multi-scale, multi-contrast descriptor set (Ferretti and Corino, 2025; Rezaeijo et al., 2025).
Figure 3.
Illustration of radiomic feature composition. A single bright lesion region-of-interest (ROI) is shown together with the corresponding first-order intensity histogram and four classes of texture matrices—gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size-zone matrix (GLSZM), and gray-level dependence matrix (GLDM)—all computed within the lesion mask.
2.5. Clinical characteristics
This study gathered demographic and baseline clinical characteristics through a neurologist’s assessment. Collected data included sex, age, and medical history, covering hypertension, diabetes, smoking, alcohol consumption, atrial fibrillation, prior strokes, and coronary heart disease. Additionally, we recorded the TOAST classification for the subtypes of acute ischemic stroke, the occlusion territory (anterior or posterior circulation), baseline mRS and National Institutes of Health Stroke Scale (NIHSS) scores, alongside Glasgow Coma Scale (GCS) scores and the Alberta Stroke Program Early CT Score (ASPECTS). Laboratory parameters included glycated hemoglobin (HbA1c), homocysteine (Hcy), low-density lipoprotein (LDL), high-density lipoprotein (HDL), total cholesterol, triglycerides, uric acid, creatinine, C-reactive protein (CRP), cystatin C (Cys-C), brain natriuretic peptide (BNP), D dimer, neutrophil, lymphocyte, the neutrophil/lymphocyte ratio (NLR), and fibrinogen. Interventional angiographic variables were documented by the treating neurointerventionalist, including the type of anesthesia (general or local), bridging therapy, number of thrombectomy passes, and the intervals from symptom onset to puncture and recanalization, as well as the extent of recanalization mTICI score. Univariate screening guided subsequent feature selection, and patient clinical outcomes were evaluated using the mRS score at 90 days post-intervention. FR was defined as an unfavorable functional outcome, indicated by an mRS score > 2 despite a technically successful intervention (mTICI ≥ 2b). We adopted this definition because mRS > 2 at 90 days after successful recanalization is commonly used in prior EVT studies to represent loss of functional independence despite angiographic success (Hussein et al., 2010; Nie et al., 2023).
2.6. Feature engineering and dimensionality reduction
We utilized univariate ANOVA F-tests (SelectKBest) to rank features based on their association with the binary outcome within the training data only. To preserve the diversity and interpretability of modalities, we retained the top 5 radiomics features from each modality for the image-only models. Within each cross-validation fold, the retained features were then standardized (zero-mean, unit variance) using parameters estimated from the corresponding training subset only (Pedregosa et al., 2011).
2.7. Model development and training
The dataset was stratified based on outcomes and randomly divided into two partitions: 70% for training and 30% for testing. We trained five classifiers: Support Vector Machine (SVM) with RBF kernel, Random Forest, Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) (Chen and Guestrin, 2016; Ke et al., 2017; Ostroumova et al., 2017; Tibshirani, 2018). For multimodal fusion models, concatenated feature sets subsequently underwent recursive feature elimination (RFE) within the same training subset before classifier fitting. To address class imbalance, RandomOverSampler/SMOTE from imbalanced-learn was then applied only to the training portion of each fold, thereby avoiding data leakage (Lemaître et al., 2017). Hyperparameters, including learning rate, maximum depth, L1/L2 regularization, and number of estimators, were optimized by grid search within the 5-fold cross-validation framework using the 70% training partition only. After model selection, the best-performing configuration was refit on the complete training partition, and the untouched 30% test set was used once for final performance evaluation (Li et al., 2022).
2.8. Ensemble multimodal strategy
We developed (i) single-modality CT radiomics models for each of the six modalities; (ii) dual-modality models by concatenating feature vectors from NCCT with one CTP map (e.g., NC + CBF, NC + CBV, NC + MTT, NC + TTP); and (iii) tri-modality models (e.g., NC + CBV + TTP). For clinical-imaging fusion, we selected five preoperative clinical variables (Cys-C, lymphocyte, NIHSS, onset-to-recanalization time, ASPECTS) using the same selection procedure and combined them with the radiomics features for joint modeling (Figure 4). We opted for early fusion (feature-level concatenation) to enhance transparency and ensured consistent preprocessing across all pipelines. To mitigate dimensionality inflation in higher-order fusions, we limited the selection to the top features from each modality and re-applied RFE within each fusion scenario (Wu et al., 2025).
Figure 4.
Top AUCs of all models and their fusion strategies. AUC, the area under the ROC curve; LGBM, Light Gradient Boosting Machine; RF, Random Forest; CAT, Categorical Boosting; XGB, Extreme Gradient Boosting.
2.9. Performance evaluation
Primary discrimination was assessed by the area under the receiver operating characteristic (ROC) curve (AUC), with results compared using DeLong’s test (DeLong et al., 1988). Additionally, we reported metrics such as accuracy, precision (positive predictive value), recall (sensitivity), specificity, and F1-score at the Youden-optimized threshold. Model calibration was assessed using calibration curves and Brier scores, while decision curve analysis quantified the net benefit across clinically relevant thresholds (Vickers and Elkin, 2006).
3. Results
In terms of clinical outcomes (mRS) for AIS patients 90 days after EVT, FR was observed in 66/101 patients (65%), with 35 patients achieving meaningful recanalization. Figure 4 illustrates top AUCs of all models. The details of the multimodal dataset and model construction are outlined below. In summary, the multimodal CT radiomics-clinical ensemble model outperformed others, while both CT radiomics and clinical models demonstrated strong predictive capabilities for FR in AIS patients undergoing EVT. Notably, XGBoost and CatBoost emerged as the most consistently effective classifiers for the complex radiomics-clinical multimodal data among the five classifiers.
3.1. Demographic and clinical characteristics
Baseline clinical, laboratory, and angiographic characteristics of patients with FR compared to those with meaningful recanalization are summarized in Table 1. In summary, the FR group was older than the meaningful recanalization group (69.56 vs. 61.03; p = 0.002) and exhibited a higher incidence of atrial fibrillation (27.3% vs. 0%; p < 0.001). Additionally, the FR group had higher median baseline mRS (5 [4, 5] vs. 4 [3, 4]; p < 0.001) and NIHSS scores (13 [8, 18] vs. 8 [6, 10]; p < 0.001), alongside lower GCS (10 [6, 14] vs. 13 [11, 15]; p = 0.038) and ASPECTS (8 [7, 8] vs. 9 [8, 9]; p < 0.001). The FR group also had higher baseline serum Cys-C (1.11 vs. 0.90; p = 0.010) and BNP levels (1139.64 vs. 861.32; p = 0.026), a lower lymphocyte (1.38 vs. 1.72; p = 0.006), and higher NLR (10.00 vs. 4.98; p = 0.025). Furthermore, the FR group experienced longer onset-to-start of endovascular treatment times (429.09 vs. 374.11 min; p = 0.048) and longer onset-to-recanalization times (518.94 vs. 429.06 min; p = 0.002). The remaining characteristics did not show significant differences between the two groups.
Table 1.
Baseline characteristics by outcome.
| Variable | Meaningful recanalization N = 351 | Futile recanalization N = 661 | p-value2 |
|---|---|---|---|
| Clinical | |||
| Sex | 0.087 | ||
| Female | 9 (26%) | 29 (44%) | |
| Male | 26 (74%) | 37 (56%) | |
| Age (years) | 61.03 ± 11.88 | 69.56 ± 13.88 | 0.002 |
| Hypertension (yes) | 68.6% | 56.1% | 0.286 |
| Diabetes (yes) | 25.7% | 34.8% | 0.379 |
| Smoking (yes) | 34.3% | 24.2% | 0.351 |
| Drinking (yes) | 8.6% | 12.1% | 0.743 |
| Atrial fibrillation (yes) | 0.0% | 27.3% | <0.001 |
| Prior stroke | >0.999 | ||
| No | 28 (80%) | 52 (80%) | |
| Yes | 7 (20%) | 13 (20%) | |
| Unknown | 0 | 1 | |
| Coronary artery disease (yes) | 5.7% | 12.1% | 0.487 |
| TOAST classification | 0.400 | ||
| Large-artery atherosclerosis | 26 (74%) | 41 (62%) | |
| Cardioembolism | 7 (20%) | 21 (32%) | |
| Stroke of other determined etiology | 2 (5.7%) | 4 (6.1%) | |
| Occlusion territory | >0.999 | ||
| Anterior circulation | 29 (83%) | 54 (82%) | |
| Posterior circulation | 6 (17%) | 12 (18%) | |
| mRS (baseline) | 4 [3, 4] | 5 [4, 5] | <0.001 |
| NIHSS (baseline) | 8 [6, 10] (3–20) | 13 [8, 18] (1–31) | <0.001 |
| Glasgow Coma Scale (GCS) | 13 [11, 15] (3–15) | 10 [6, 14] (3–15) | 0.038 |
| ASPECTS | 9 [8, 9] (7–9) | 8 [7, 8] (6–9) | <0.001 |
| Laboratory | |||
| Glycated hemoglobin (HbA1c) | 6.22 ± 1.39 | 6.46 ± 1.84 | 0.828 |
| Homocysteine (Hcy) | 13.74 ± 9.43 | 14.33 ± 7.98 | 0.516 |
| Low-density lipoprotein (LDL) | 2.68 ± 0.73 | 3.00 ± 1.14 | 0.261 |
| High-density lipoprotein (HDL) | 0.98 ± 0.23 | 1.05 ± 0.27 | 0.237 |
| Total cholesterol | 4.27 ± 0.90 | 4.62 ± 1.36 | 0.418 |
| Triglycerides | 1.32 ± 0.57 | 1.43 ± 1.06 | 0.814 |
| Uric acid | 376.91 ± 109.18 | 391.74 ± 136.03 | 0.633 |
| Creatinine | 85.40 ± 51.75 | 88.20 ± 29.42 | 0.099 |
| C-reactive protein (CRP) | 11.90 ± 14.01 | 23.20 ± 36.14 | 0.128 |
| Cystatin C | 0.90 ± 0.51 | 1.11 ± 0.58 | 0.010 |
| Brain natriuretic peptide (BNP) | 861.32 ± 3,097.90 | 1,139.64 ± 1,870.95 | 0.026 |
| D-dimer | 2,020.20 ± 4,565.42 | 1,914.58 ± 2,686.48 | 0.237 |
| Neutrophils | 7.33 ± 2.14 | 8.38 ± 4.18 | 0.348 |
| Lymphocytes | 1.72 ± 0.71 | 1.38 ± 0.84 | 0.006 |
| Neutrophil-to-lymphocyte ratio (NLR) | 4.98 ± 2.72 | 10.00 ± 10.82 | 0.025 |
| Fibrinogen | 3.44 ± 1.28 | 3.58 ± 1.13 | 0.309 |
| Angiographic | |||
| Anesthesia type | >0.999 | ||
| Local | 30 (86%) | 56 (85%) | |
| General | 5 (14%) | 10 (15%) | |
| Bridging therapy (Yes) | 22.9% | 16.7% | 0.593 |
| Onset-to-puncture, min | 374.11 ± 105.51 | 429.09 ± 136.15 | 0.048 |
| Onset-to-recanalization, min | 429.06 ± 110.56 | 518.94 ± 136.68 | 0.002 |
| Number of retrieval attempts | 0.061 | ||
| 1 | 22 (63%) | 29 (44%) | |
| 2 | 12 (34%) | 23 (35%) | |
| 3 | 1 (2.9%) | 12 (18%) | |
| 4 | 0 (0%) | 2 (3.0%) | |
| mTICI (post-recanalization) | 0.141 | ||
| 2b | 5 (14%) | 19 (29%) | |
| 3 | 30 (86%) | 47 (71%) | |
n (%), Mean ± SD.
Pearson’s Chi-squared test, Fisher’s exact test, Wilcoxon rank sum test. Bold values mean p < 0.05, and was considered statistically significant.
mRS, modified Rankin Scale; NIHSS, National Institutes of Health Stroke Scale; ASPECTS, Alberta Stroke Program Early CT Score; mTICI, modified Thrombolysis in Cerebral Infarction.
3.2. Single-modality CT radiomics models
Among six single-modality CT models, perfusion (CTP) maps outperformed structural CT (NCCT) for outcome discrimination, while CECT consistently underperformed. For NCCT, the most effective model (LightGBM) achieved an AUC of 0.782, with an accuracy of 0.774, sensitivity of 0.850, specificity of 0.636, and F1 score of 0.829. CECT produced AUCs < 0.60 in all algorithms (best: Random Forest AUC = 0.564). Given its uniformly poor performance, CECT was not considered in subsequent imaging-only fusions. In terms of CTP, the modalities incorporating CBF and MTT exhibited the highest efficiency. The CBF model (CatBoost) achieved an AUC of 0.832 (accuracy = 0.710, sensitivity = 0.800, specificity = 0.545, F1 = 0.780), while the XGBoost model using MTT attained an AUC of 0.855 (accuracy = 0.806, sensitivity = 0.850, specificity = 0.727, F1 = 0.850) (Table 2; Figure 5).
Table 2.
Results of single-modality CT radiomics models.
| Modality | Best algorithm | Accuracy | Precision | Recall | Specificity | F1 score | AUC |
|---|---|---|---|---|---|---|---|
| NCCT | LightGBM | 0.774 | 0.810 | 0.850 | 0.636 | 0.829 | 0.782 |
| CECT | Random forest | 0.548 | 0.667 | 0.600 | 0.455 | 0.632 | 0.564 |
| CBF | CatBoost | 0.710 | 0.762 | 0.800 | 0.545 | 0.780 | 0.832 |
| CBV | XGBoost | 0.742 | 0.773 | 0.850 | 0.545 | 0.810 | 0.795 |
| MTT | XGBoost | 0.806 | 0.850 | 0.850 | 0.727 | 0.85 | 0.855 |
| TTP | Random forest | 0.710 | 0.739 | 0.850 | 0.455 | 0.791 | 0.782 |
Figure 5.
The performance of single-modality CT radiomics models.
3.3. Multi-modalities CT radiomics models
Fusing modalities improved performance relative to single-modality models, with the magnitude dependent on the complementarity of paired or triplet inputs. Among dual-modality models, NCCT+CBF (LightGBM) performed best with AUC = 0.895 (accuracy = 0.774, sensitivity = 0.900, specificity = 0.545, F1 = 0.837). Tri-modality fusion yielded additional gains in selected combinations. The top tri-modality model was NCCT+CBF + MTT (LightGBM) with AUC = 0.900, accuracy = 0.839, sensitivity = 0.900, specificity = 0.727, and F1 = 0.878 (Table 3; Figures 6a,b; Supplementary Figures 1–8; Supplementary Tables 1–8).
Table 3.
Results of multi-modalities CT radiomics models.
| Combination | Best algorithm | Accuracy | Precision | Recall | Specificity | F1 score | AUC |
|---|---|---|---|---|---|---|---|
| NCCT + CBF | LightGBM | 0.774 | 0.783 | 0.900 | 0.545 | 0.837 | 0.895 |
| NCCT + CBV | XGBoost | 0.774 | 0.810 | 0.850 | 0.636 | 0.829 | 0.832 |
| NCCT + MTT | LightGBM | 0.742 | 0.833 | 0.750 | 0.727 | 0.789 | 0.855 |
| NCCT + TTP | Random forest | 0.677 | 0.708 | 0.850 | 0.364 | 0.773 | 0.836 |
| NCCT + CBF + MTT | LightGBM | 0.839 | 0.857 | 0.900 | 0.727 | 0.878 | 0.900 |
| NCCT + CBV + MTT | LightGBM | 0.871 | 0.864 | 0.950 | 0.727 | 0.905 | 0.886 |
| NCCT + CBF + TTP | LightGBM | 0.710 | 0.762 | 0.800 | 0.545 | 0.780 | 0.855 |
| NCCT + CBV + TTP | Random forest | 0.806 | 0.792 | 0.950 | 0.545 | 0.864 | 0.864 |
Figure 6.
The best performance of Multi-modalities CT radiomics models and multimodal radiomics-clinical ensemble models, as well as the results of clinical models. (a) NCCT + CBF (LightGBM) performed best among dual-modality models; (b) NCCT + CBF + MTT (LightGBM) performed best among tri-modality models; (c) The performance of clinical models; (d) Clinical + MTT (Random Forest) performed best among clinical + single CT modality; (e) Clinical + NCCT + CBV (CatBoost) performed best among clinical + two CT modalities; (f) Clinical + NCCT + CBV + TTP (Random Forest) performed best among clinical + three CT modalities.
3.4. Clinical models
Using five clinically selected variables (Cys-C, lymphocyte, baseline NIHSS, onset-to-recanalization, ASPECTS), the clinical-only models demonstrated moderate performance, with the highest AUC reaching 0.800 (CatBoost). This model achieved an accuracy of 0.742, sensitivity of 0.900, specificity of 0.455, and an F1 score of 0.818. The Random Forest model showed comparable results with an AUC of 0.798 (Table 4; Figure 6c).
Table 4.
Results of clinical models.
| Algorithm | Accuracy | Precision | Recall | Specificity | F1 Score | AUC |
|---|---|---|---|---|---|---|
| SVM | 0.548 | 0.800 | 0.400 | 0.818 | 0.533 | 0.600 |
| Random forest | 0.742 | 0.773 | 0.850 | 0.545 | 0.810 | 0.798 |
| LightGBM | 0.710 | 0.739 | 0.850 | 0.455 | 0.791 | 0.757 |
| XGBoost | 0.710 | 0.739 | 0.850 | 0.455 | 0.791 | 0.759 |
3.5. Multimodal radiomics-clinical ensemble models
The integration of clinical variables with radiomics consistently enhanced model performance compared to either radiomics-only or clinical-only approaches. For the clinical plus single radiomics modality, the optimal results were observed with the clinical + NCCT (CatBoost) model, which attained an AUC of 0.886 (accuracy = 0.839, sensitivity = 0.950, specificity = 0.636, F1 = 0.884). This outcome was superior to both the single clinical model (AUC = 0.800) and the NCCT-only model (AUC = 0.782). Following the combination of clinical and CTP modalities, the AUC for the Random Forest model using clinical + MTT rose to 0.889, marking an improvement over the single clinical model (AUC = 0.800) and the MTT-only model (AUC = 0.855). Notably, the ensemble model integrating clinical, NCCT, and CTP modalities–the clinical + NCCT + CBV model (CatBoost) –achieved the highest performance in the study, with an AUC of 0.918, accuracy of 0.806, sensitivity of 0.950, specificity of 0.545, and F1 score of 0.864. Additional pairings also improved over their radiomics-only counterparts, e.g., clinical+NCCT+TTP (Random Forest) AUC = 0.907 and clinical+NCCT+CBF (Random Forest) model reached an AUC of 0.855. However, when combining clinical variables with three imaging modalities, the results were mixed and did not consistently surpass the best dual-fusion models (Table 5; Figures 6d–f; Supplementary Figures 9–22; Supplementary Tables 9–22).
Table 5.
Results of multimodal radiomics-clinical ensemble models.
| Combination | Best algorithm | Accuracy | Precision | Recall | Specificity | F1 Score | AUC |
|---|---|---|---|---|---|---|---|
| Clinical + NCCT | CatBoost | 0.839 | 0.826 | 0.950 | 0.636 | 0.884 | 0.886 |
| Clinical + CECT | Random forest | 0.806 | 0.850 | 0.850 | 0.727 | 0.850 | 0.830 |
| Clinical + CBF | CatBoost | 0.742 | 0.750 | 0.900 | 0.455 | 0.818 | 0.841 |
| Clinical + CBV | CatBoost | 0.806 | 0.818 | 0.900 | 0.636 | 0.857 | 0.882 |
| Clinical + MTT | Random forest | 0.742 | 0.750 | 0.900 | 0.455 | 0.818 | 0.889 |
| Clinical + TTP | CatBoost | 0.742 | 0.750 | 0.900 | 0.455 | 0.818 | 0.873 |
| Clinical + NCCT + CBF | Random forest | 0.774 | 0.783 | 0.900 | 0.545 | 0.837 | 0.855 |
| Clinical + NCCT + CBV | CatBoost | 0.806 | 0.792 | 0.950 | 0.545 | 0.864 | 0.918 |
| Clinical + NCCT + MTT | Random forest | 0.774 | 0.760 | 0.950 | 0.455 | 0.844 | 0.848 |
| Clinical + NCCT + TTP | Random forest | 0.774 | 0.760 | 0.950 | 0.455 | 0.844 | 0.907 |
| Clinical + NCCT + CBF + MTT | Random forest | 0.774 | 0.760 | 0.950 | 0.455 | 0.844 | 0.905 |
| Clinical + NCCT + CBV + MTT | LightGBM | 0.742 | 0.731 | 0.950 | 0.364 | 0.826 | 0.855 |
| Clinical + NCCT + CBF + TTP | Random forest | 0.806 | 0.818 | 0.900 | 0.636 | 0.857 | 0.891 |
| Clinical + NCCT + CBV + TTP | Random forest | 0.774 | 0.760 | 0.950 | 0.455 | 0.844 | 0.907 |
3.6. Classifier performance comparison
We assessed five distinct machine learning classifiers (SVM, Random Forest, LightGBM, XGBoost, and CatBoost) to identify the most effective model for our prediction task. Overall, ensemble tree-based models, particularly the gradient boosting algorithms, exhibited a performance advantage over SVM and standard Random Forest classifiers. XGBoost and CatBoost consistently achieved the highest AUCs across most feature sets. The Random Forest classifier performed adequately but lagged behind the boosting models. The SVM classifier displayed more variability in its results. LightGBM remained consistently competitive, often finishing within a narrow margin of the leading models and occasionally best-in-class for select dual- and tri-modality fusions (e.g., NCCT+CBF, NCCT+CBF + MTT)–but on average trailed XGBoost/CatBoost by a small margin (Supplementary Figures 1–22; Supplementary Tables 1–22).
4. Discussion
FR presents a significant challenge in the context of EVT and there is a lack of reliable predictive models for assessing treatment outcomes in acute ischemic stroke. In this study, we employ machine learning techniques to combine multimodal CT radiomics, clinical, laboratory, and angiographic data for accurate prediction of FR. Ultimately, our multimodal CT radiomics-clinical ensemble machine learning model with the CatBoost algorithm demonstrates excellent performance, achieving an AUC of 0.918.
Among single CT modality models, CTP modalities exhibited the best performance, followed by NCCT, while CECT underperformed. The superior performance of CTP compared to NCCT can be attributed to its ability to provide comprehensive hemodynamic information regarding the ischemic brain tissues (Saver et al., 2015). Brugnara et al. (2020) suggested that the predictive value of CTP for FR is limited, potentially because their research focused solely on the infarct core and ischemic penumbra, rather than conducting a comprehensive radiomics analysis of individual CTP parameters. The limited contrast resolution of enhanced scans may also hinder feature discrimination in CECT. Utilizing an anatomy-function complementarity strategy, the performance of dual-modality (NCCT+CTP) models was enhanced. Furthermore, incorporating a third modality generally improved discrimination when it contributed non-redundant hemodynamic information. Among tri-modality CT radiomics, the optimal combined model was NCCT + CBF + MTT, achieving an AUC of 0.900, with the most significant gains observed when pairing NCCT with both a volume (CBV) and a time (MTT/TTP) parameter. This finding indicates that the integrated analysis of structural maps along with cerebral blood volume and hemodynamic information over time is crucial for prognostic predictions.
In the preliminary univariate analysis of clinical variables, age, baseline NIHSS, Cys-C, the time from onset to puncture and other factors were significantly associated with functional outcomes at 90 days. Most predictors align with previous studies, though some discrepancies may arise from differences in research subjects and sample sizes. Notably, consistent findings across studies (including ours) reveal that older age, higher preoperative mRS and NIHSS scores, elevated NLR, lower ASPECTS and GCS scores, and prolonged time from onset to puncture and recanalization are associated with FR (Nie et al., 2023; D’Anna et al., 2024; Deng et al., 2022; Ni et al., 2023; Shen et al., 2023). These associations are also clinically plausible. Older age and higher baseline disability may reflect reduced physiological and neurological reserve, while higher NIHSS and lower GCS are markers of more severe initial neurological injury. Lower ASPECTS generally indicates a larger extent of early ischemic change, and elevated NLR may reflect a stronger systemic inflammatory response that contributes to secondary injury. In addition, longer treatment delays may increase irreversible tissue damage before reperfusion is achieved, thereby increasing the likelihood of poor functional recovery despite technically successful recanalization. Overall, the performance of clinical models lagged behind that of CT radiomics models. This discrepancy may be due to clinical variables being more susceptible to variability and possessing lower specificity. Among the clinical models, CatBoost and Random Forest demonstrated the best comprehensive performance, with AUC of 0.800 and 0.798, respectively. This suggests that these machine learning algorithms are well-suited as prognostic prediction tools based on clinical information, providing a reliable baseline for subsequent ensemble modeling with CT radiomics data.
When integrating clinical indicators with each CT radiomics modality for modeling, the ensemble model generally outperformed both the single clinical and CT radiomics models, underscoring the advantages of ensemble modeling. Remarkably, the clinical + NCCT + CBV ensemble model using CatBoost achieved an AUC of 0.918, surpassing any single modality or other combined models. This highlights the potential of the multimodal omics integration strategy in predicting treatment efficacy after endovascular interventions, thereby supporting the pursuit of more precise individualized treatment. However, results for the clinical + three CT modalities results were mixed and did not consistently surpass the best tri-fusion models. Consequently, integrating more modalities does not always guarantee superior outcomes. Given the inherent correlations among CTP parameters, adding more modalities generally improves performance when the added maps provide non-redundant hemodynamic information. Collectively, fusion models centered on NCCT + CBV (with or without a time parameter) and enhanced by clinical variables yielded the most robust discrimination.
In this study, gradient boosting algorithms (especially XGBoost and CatBoost) showed a performance advantage over SVM and standard Random Forest across most feature sets. LightGBM consistently demonstrated competitive performance, often within a narrow margin of the top performers and occasionally excelling in selected dual- and tri-modality fusions. This finding aligns with previous studies on stroke outcome prediction (Mbarek et al., 2024). The superior performance of boosted trees can be attributed to their capability to model complex non-linear feature interactions and manage high-dimensional data with inherent regularization (shrinkage and tree pruning), making them well-suited for radiomics features. While the Random Forest classifier performed reasonably well, it lagged behind the boosting models, likely due to its inability to focus on the most informative features as effectively as gradient boosting, which iteratively minimizes residual errors. The SVM classifier displayed more variable results in our overall comparison. SVM tended to be more sensitive to the curse of dimensionality. Its performance was slightly lower with full feature sets and required careful tuning of hyperparameters (kernel parameters and regularization) to prevent overfitting. In contrast, tree-based models managed large feature pools more robustly by automatically selecting splits on the most predictive features. LightGBM utilized histogram-based leaf-wise growth with built-in regularization. These findings highlight the importance of algorithm selection in integrated multimodal studies, as ensemble learning methods can unlock more prognostic signals within the data, while simpler classifiers may underfit or overfit if not matched to the feature characteristics. Thus, fusing multiple classifiers or conducting a rigorous model selection process is recommended to ensure optimal predictive performance for a given radiomics application.
Our study has several limitations. First, the relatively limited sample size and retrospective single-center design may restrict the generalizability of the models, because institutional differences in imaging workflow, clinical decision-making, and laboratory testing may affect model applicability in other settings. Further prospective studies with larger multicenter samples are needed to validate these findings externally. Second, postoperative CT and clinical data were not included in this study. Integrating preoperative and postoperative data may further enhance predictive performance. Third, our analysis combined AIS with anterior and posterior circulation occlusions, because these subgroups may differ in pathophysiology, treatment response, and clinical prognosis, this combined analysis may have introduced heterogeneity. Future studies with larger cohorts should analyze these populations separately.
5. Conclusion
In conclusion, the multimodal CT radiomics-clinical ensemble model effectively predicts FR following EVT for AIS using gradient boosting algorithms. The ensemble model, when appropriately tailored with relevant variables and algorithms, has the potential to facilitate personalized treatment and improve clinical outcomes for AIS.
Acknowledgments
The authors thank all study participants for contributing.
Funding Statement
The author(s) declared that financial support was received for this work and/or its publication. This work was supported by the Basic and Applied Basic Research Foundation of Guangdong (2023A1515220066 and 2020A1515110165), the Funding by Science and Technology Projects in Guangzhou (2023A03J0959), and Natural Science Foundation of Tibet Autonomous Region (XZ2024ZR-ZY089 (Z)).
Footnotes
Edited by: Jin Lei, Huazhong University of Science and Technology, China
Reviewed by: Xiaowen Xu, Tongji University, China
Mingze Xu, Peking University, China
Data availability statement
The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.
Ethics statement
The studies involving humans were approved by Ethics Committee of the Guangzhou First People’s Hospital. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because written consent was not required for this study because it used anonymized retrospective data.
Author contributions
ZW: Writing – review & editing, Supervision, Writing – original draft, Conceptualization, Validation. YG: Writing – review & editing, Writing – original draft, Data curation, Conceptualization. PX: Writing – review & editing, Methodology, Data curation, Writing – original draft. DW: Formal analysis, Writing – original draft, Methodology, Conceptualization. WL: Resources, Formal analysis, Writing – original draft. HH: Writing – original draft, Funding acquisition, Resources, Project administration. WD: Project administration, Writing – review & editing, Methodology, Investigation. HX: Project administration, Validation, Supervision, Writing – review & editing. XW: Validation, Writing – review & editing, Visualization, Supervision. XL: Visualization, Funding acquisition, Validation, Conceptualization, Writing – review & editing, Supervision.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnins.2026.1838675/full#supplementary-material
References
- Brugnara G., Neuberger U., Mahmutoglu M. A., Foltyn M., Herweh C., Nagel S., et al. (2020). Multimodal predictive modeling of endovascular treatment outcome for acute ischemic stroke using machine-learning. Stroke 51, 3541–3551. doi: 10.1161/STROKEAHA.120.030287, [DOI] [PubMed] [Google Scholar]
- Chen T., Guestrin C., (2016) XGBoost: A Scalable Tree Boosting System. Fort Worth, TX, USA: JKL International LLC. [Google Scholar]
- Chen Q., Xia T., Zhang M., Xia N., Liu J., Yang Y. (2021). Radiomics in stroke neuroimaging: techniques, applications, and challenges. Aging Dis. 12, 143–154. doi: 10.14336/AD.2020.0421, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Choy G., Khalilzadeh O., Michalski M., Do S., Samir A. E., Pianykh O. S., et al. (2018). Current applications and future impact of machine learning in radiology. Radiology 288, 318–328. doi: 10.1148/radiol.2018171820 [DOI] [PMC free article] [PubMed] [Google Scholar]
- D’Anna L., Merlino G., Romoli M., Zhang L., del Regno C., Aggour M., et al. (2024). Predictors of futile recanalization in nonagenarians treated with mechanical thrombectomy: a multi-center observational study. J. Neurol. 271, 4925–4932. doi: 10.1007/s00415-024-12428-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Da Ros V., Cavallo A., Di Donna C., D'Onofrio A., Trulli M., Di Candia S., et al. (2024). Ensemble machine learning to predict futile recanalization after mechanical thrombectomy based on non-contrast CT imaging. J. Stroke Cerebrovasc. Dis. 33:107890. doi: 10.1016/j.jstrokecerebrovasdis.2024.107890 [DOI] [PubMed] [Google Scholar]
- DeLong E. R., DeLong D. M., Clarke-Pearson D. L. (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 44, 837–845. doi: 10.2307/2531595 [DOI] [PubMed] [Google Scholar]
- Deng G., Xiao J., Yu H., Chen M., Shang K., Qin C., et al. (2022). Predictors of futile recanalization after endovascular treatment in acute ischemic stroke: a meta-analysis. J NeuroIntervent Surg 14, 881–885. doi: 10.1136/neurintsurg-2021-017963, [DOI] [PubMed] [Google Scholar]
- Ferretti M., Corino V. D. A. (2025). Integrating radiomic and 3D autoencoder-based features for non-small cell lung Cancer survival analysis. Comput. Methods Prog. Biomed. 258:108496. doi: 10.1016/j.cmpb.2024.108496, [DOI] [PubMed] [Google Scholar]
- Goyal M., Menon B. K., van Zwam W. H., Dippel D. W., Mitchell P. J., Demchuk A. M., et al. (2016). Endovascular thrombectomy after large-vessel ischaemic stroke: a meta-analysis of individual patient data from five randomised trials. Lancet 387, 1723–1731. doi: 10.1016/S0140-6736(16)00163-X, [DOI] [PubMed] [Google Scholar]
- Happi Ngankou E., Gory B., Marnat G., Richard S., Bourcier R., Sibon I., et al. (2021). Thrombectomy complications in large vessel occlusions: incidence, predictors, and clinical impact in the ETIS registry. Stroke 52, e764–e768. doi: 10.1161/STROKEAHA.121.034865, [DOI] [PubMed] [Google Scholar]
- Hussein H. M., Georgiadis A. L., Vazquez G., Miley J. T., Memon M. Z., Mohammad Y. M., et al. (2010). Occurrence and predictors of futile recanalization following endovascular treatment among patients with acute ischemic stroke: a multicenter study. AJNR Am. J. Neuroradiol. 31, 454–458. doi: 10.3174/ajnr.A2006, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Jenkinson M., Bannister P., Brady M., Smith S. (2002). Improved optimization for the robust and accurate linear registration and motion correction of brain images. NeuroImage 17, 825–841. doi: 10.1006/nimg.2002.1132, [DOI] [PubMed] [Google Scholar]
- Jia B., Ren Z., Mokin M., Burgin W. S., Bauer C. T., Fiehler J., et al. (2021). Current status of endovascular treatment for acute large vessel occlusion in China: a real-world Nationwide registry. Stroke 52, 1203–1212. doi: 10.1161/STROKEAHA.120.031869, [DOI] [PubMed] [Google Scholar]
- Ke G., Meng Q., Finley T., Wang T., Chen W., Ma W., et al. (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree NIPS'17: Proceedings of the 31st International Conference on Neural Information Processing Systems, [Google Scholar]
- Lemaître G., Nogueira F., Aridas C. K. (2017). Imbalanced-learn: a python toolbox to tackle the curse of imbalanced datasets in machine learning. J Mach Learn Res 18, 559–563. doi: 10.48550/arXiv.1609.06570 [DOI] [Google Scholar]
- Li Y., Liu Y., Hong Z., Wang Y., Lu X. (2022). Combining machine learning with radiomics features in predicting outcomes after mechanical thrombectomy in patients with acute ischemic stroke. Comput. Methods Prog. Biomed. 225:107093. doi: 10.1016/j.cmpb.2022.107093, [DOI] [PubMed] [Google Scholar]
- Lin Y., Schulze V., Brockmeyer M., Parco C., Karathanos A., Heinen Y., et al. (2019). Endovascular Thrombectomy as a means to improve survival in acute ischemic stroke: a Meta-analysis. JAMA Neurol. 76, 850–854. doi: 10.1001/jamaneurol.2019.0525, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Luo Y., Sun X., Kong X., Tong X., Xi F., Mao Y., et al. (2023). A DWI-based radiomics-clinical machine learning model to preoperatively predict the futile recanalization after endovascular treatment of acute basilar artery occlusion patients. Eur. J. Radiol. 161:110731. doi: 10.1016/j.ejrad.2023.110731, [DOI] [PubMed] [Google Scholar]
- Mbarek L., Chen S., Jin A., Pan Y., Meng X., Yang X., et al. (2024). Predicting 3-month poor functional outcomes of acute ischemic stroke in young patients using machine learning. Eur. J. Med. Res. 29:494. doi: 10.1186/s40001-024-02056-3, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ni H., Wang B., Hang Y., Liu S., Jia Z. Y., Shi H. B., et al. (2023). Predictors of futile recanalization in patients with intracranial atherosclerosis-related stroke undergoing endovascular treatment. World Neurosurg. 171, e752–e759. doi: 10.1016/j.wneu.2022.12.101, [DOI] [PubMed] [Google Scholar]
- Nie X., Leng X., Miao Z., Fisher M., Liu L. (2023). Clinically ineffective reperfusion after endovascular therapy in acute ischemic stroke. Stroke 54, 873–881. doi: 10.1161/STROKEAHA.122.038466 [DOI] [PubMed] [Google Scholar]
- Ostroumova L., Gusev G., Vorobev A., Dorogush A. V., Gulin A. (2017) CatBoost: Unbiased Boosting with Categorical features. Neural Inform. Process. Syst. 2017 6639–6649 [Google Scholar]
- Pedregosa F., Varoquaux G., Gramfort A., Rakuten M. (2011). Scikit-Learn: Machine Learning in Python. J Mach Learn Res 12, 2825–2830. doi: 10.48550/arXiv.1201.0490 [DOI] [Google Scholar]
- Pisani N., Destito M., Ricciardi C., Pellecchia M. T., Cesarelli M., Esposito F., et al. (2025). Repeatability of radiomic features from brain T1-W MRI after image intensity normalization: implications for longitudinal studies on structural neurodegeneration. Comput. Methods Prog. Biomed. 265:108738. doi: 10.1016/j.cmpb.2025.108738, [DOI] [PubMed] [Google Scholar]
- Powers W. J., Rabinstein A. A., Ackerson T., Adeoye O. M., Bambakidis N. C., Becker K., et al. (2019). Guidelines for the early Management of Patients with Acute Ischemic Stroke: 2019 update to the 2018 guidelines for the early Management of Acute Ischemic Stroke: a guideline for healthcare professionals from the American Heart Association/American Stroke Association. Stroke 50, e344–e418. doi: 10.1161/STR.0000000000000211, [DOI] [PubMed] [Google Scholar]
- Rezaeijo S. M., Eftekhar A., Rouhi S., Keshavarzi B., Mohammadi Z., Firouzabad L. A., et al. (2025). Neighboring tissues as diagnostic windows: neighborhood effects in radiomic detection of pancreatic ductal adenocarcinoma. Comput. Methods Prog. Biomed. 272:109056. doi: 10.1016/j.cmpb.2025.109056, [DOI] [PubMed] [Google Scholar]
- Saver J. L., Goyal M., Bonafe A., Diener H. C., Levy E. I., Pereira V. M., et al. (2015). Stent-retriever thrombectomy after intravenous t-PA vs. t-PA alone in stroke. N. Engl. J. Med. 372, 2285–2295. doi: 10.1056/NEJMoa1415061, [DOI] [PubMed] [Google Scholar]
- Shen H., Killingsworth M. C., Bhaskar S. M. M. (2023). Comprehensive Meta-analysis of futile recanalization in acute ischemic stroke patients undergoing endovascular Thrombectomy: prevalence, factors, and clinical outcomes. Life 13:965. doi: 10.3390/life13101965, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Su M., Zhou Y., Chen Z., Pu M., Li Z., du H., et al. (2021). Cystatin C predicts futile recanalization in patients with acute ischemic stroke after endovascular treatment. J. Neurol. 269, 966–972. doi: 10.1007/s00415-021-10680-w [DOI] [PubMed] [Google Scholar]
- Tibshirani R. (2018). Regression shrinkage and selection via the lasso. J Royal Stat Soc Series B 58, 267–288. doi: 10.1111/j.2517-6161.1996.tb02080.x [DOI] [Google Scholar]
- van Griethuysen J. J. M., Fedorov A., Parmar C., Hosny A., Aucoin N., Narayan V., et al. (2017). Computational Radiomics system to decode the radiographic phenotype. Cancer Res. 77, e104–e107. doi: 10.1158/0008-5472.CAN-17-0339, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Varghese B. A., Cen S. Y., Hwang D. H., Duddalwar V. A. (2019). Texture analysis of imaging: what radiologists need to know. AJR Am. J. Roentgenol. 212, 520–528. doi: 10.2214/AJR.18.20624, [DOI] [PubMed] [Google Scholar]
- Vickers A. J., Elkin E. B. (2006). Decision curve analysis: a novel method for evaluating prediction models. Med. Decis. Mak. 26, 565–574. doi: 10.1177/0272989X06295361, [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wu N., Lu W., Xu M. (2025). Multimodal radiomics based on lesion connectome predicts stroke prognosis. Comput. Methods Prog. Biomed. 263:108701. doi: 10.1016/j.cmpb.2025.108701, [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.