Abstract
Background
Acute ischemic stroke (AIS) is a leading cause of mortality and morbidity worldwide. This study explores the potential of using carotid duplex ultrasound (CDU) radiomics to predict outcomes in AIS patients, enhancing prognostic assessments through advanced imaging technologies.
Methods
We enrolled 105 AIS patients at the Stroke Center of the First Affiliated Hospital of Soochow University. CDU images were obtained and processed using artificial intelligence to extract 1,477 radiomic features. Key features were identified and used to develop a predictive model for patient outcomes 3 months post-stroke.
Results
The developed predictive model demonstrated high accuracy and clinical utility, confirmed by decision curve analysis. Inter-observer reliability was excellent, with a Cohen’s kappa coefficient of 0.93, indicating consistent assessments of plaque vulnerability across different observers.
Conclusion
Carotid duplex ultrasound radiomics significantly enhances the ability to predict adverse outcomes in AIS, offering a precision-based approach that supports personalized management strategies. It could improve clinical outcomes by allowing targeted interventions based on specific radiomic profiles.
Keywords: acute ischemic stroke, carotid atherosclerosis, decision curve analysis, deep learning, prognostic assessment, ultrasound radiomics
Background
Acute ischemic stroke (AIS) accounts for approximately 80% of all stroke cases (1) and is associated with high rates of disability, mortality, and recurrence (2–4). Accurate prognostic prediction, risk stratification, and effective management are essential for improving secondary stroke prevention. While conventional tools such as the Essen Stroke Risk Score, ABCD2 score, and other prognostic scales remain widely used, the rapid growth of healthcare big data has introduced new opportunities (5). The expanding volume and complexity of medical data have increasingly enabled the application of imaging omics in disease prediction, prognosis, diagnosis, and health management (6–8).
Traditional risk score models in domestic and international AIS studies have made some progress in outcome prediction. However, their accuracy remains limited due to reliance on basic indicators such as age and medical history (8). To address this limitation, our study integrates carotid duplex ultrasound (CDU)-based radiomics with clinical parameters to enhance prognostic precision. Although several models based on clinical factors have been proposed, comprehensive studies incorporating CDU imaging features are still limited (9, 10).
AIS is a multifactorial disease in which immune-inflammatory mechanisms play a critical role (11–13). The systemic immune-inflammation index (SII), a novel marker derived from standard blood counts, offers an integrated view of immune and inflammatory status. While its prognostic value has been established in cardiovascular diseases and malignancies (14–17), its role in AIS remains underexplored. Moreover, accumulating evidence suggests that carotid stenosis severity alone is insufficient to predict stroke risk (18). Instead, plaque vulnerability has emerged as a key determinant of ischemic events. Radiomics has enabled the extraction of subtle image features beyond visual assessment, positioning itself as a valuable tool in AIS research (19, 20). Previous studies have also highlighted the association between carotid plaque imaging characteristics and AIS outcomes (21).
Although high-field MRI (3T/7T)-based radiomics allows high-resolution analysis of carotid plaque composition (22), its clinical applicability remains limited due to high costs and restricted equipment availability. In contrast, CDU offers advantages such as low cost, real-time imaging, and operational convenience, making it more suitable for widespread use in primary care. However, conventional CDU assessments are often subjective and lack standardized quantitative metrics, limiting their role in precision medicine. To address these limitations, we propose an integrated approach that combines radiomic features extracted from CDU with the SI, a blood-derived marker reflecting systemic inflammatory status. Given the time-sensitive nature of AIS management (13), CDU’s rapid acquisition and repeatability make it ideal for real-time monitoring of plaque evolution, potentially compensating for the inefficiencies and limitations of MRI in acute settings.
In parallel, machine learning has shown strong potential in acute disease prediction using routine laboratory data, particularly in acute kidney injury (AKI) (23–25). Similar strategies have been applied to AIS prognostication using common biomarkers such as albumin and urinary creatinine, achieving promising accuracy. Building on this, our study aims to evaluate the feasibility and accuracy of CDU-based radiomic features in identifying plaque vulnerability and to construct a multimodal prediction model by integrating radiomics with SII and other clinical parameters. We further assess its adaptability and translational potential across different healthcare settings. Ultimately, we seek to develop a rapid, scalable, and quantifiable risk stratification tool to support individualized decision-making and improve functional outcomes in AIS patients.
Methods
Study population
From January 2018 to December 2020, we prospectively enrolled 105 patients diagnosed with acute ischemic stroke (AIS) associated with carotid atherosclerotic plaques at the Stroke Center of the First Affiliated Hospital of Soochow University. CDU images were obtained for all participants (Figure 1).
Figure 1.
The process of patient inclusion and exclusion.
Inclusion criteria were as follows: (1) AIS confirmed with atherosclerotic abnormalities in the extracranial segments of the internal or common carotid arteries; (2) age ≥18 years; (3) admission within 3 days of symptom onset; (4) receipt of standard medical treatment; (5) CDU performed within 24 h of admission; (6) availability of three-month follow-up data; and (7) brain MRI with diffusion-weighted imaging (DWI) or CT perfusion performed upon admission with image quality sufficient for infarct volume quantification. Exclusion criteria included: (1) diagnosis of hemorrhagic cerebrovascular disease or intracranial arterial stenosis; (2) atrial fibrillation or other arrhythmias suggestive of cardioembolic stroke; (3) severe cardiovascular, hematologic, hepatic, or renal diseases; (4) concomitant acute infection, sepsis, malignancy, or autoimmune disease; (5) ongoing treatment with immunosuppressants, corticosteroids, or cytotoxic drugs; (6) history of thrombolysis, thrombectomy, carotid endarterectomy, or stenting; (7) absence of carotid plaques or complete occlusion of either the internal or common carotid artery; (8) baseline modified Rankin Scale (mRS) score >2; and (9) incomplete clinical, laboratory, or imaging data, including suboptimal image clarity or missing follow-up.
Patients were categorized into anterior and posterior circulation groups based on TOAST classification and neuroimaging to reduce heterogeneity. The anterior circulation group was defined by infarctions in the MCA or ACA territories on DWI or an ASPECTS score ≤7 on CT; the posterior circulation group was defined by infarctions in the PCA or vertebrobasilar territories on DWI or a PC-ASPECTS score ≤8 on CT. Infarct volumes were manually segmented using ITK-SNAP software on DWI images. CT images were assessed using the Alberta Stroke Program Early CT Score (ASPECTS). Baseline clinical parameters were recorded within 24 h of admission. The systemic immune-inflammation index (SII) was calculated using the formula: SII = Platelet count × Neutrophil count/Lymphocyte count. The primary outcome was the functional status 3 months post-stroke, assessed by trained professionals via standardized telephone interviews using the modified Rankin Scale (mRS). Patients were classified into favorable (mRS 0–2) or unfavorable (mRS ≥ 3) outcome groups. Patients with baseline mRS > 2 were excluded to ensure uniformity in baseline functional status.
Additional variables, including age, sex, baseline NIHSS score, and treatment modalities, were collected and adjusted for in multivariate analyses. To control for potential confounding, stratified analyses were conducted based on age (≤65 vs. >65 years), NIHSS score (≤5 vs. >5), and SII (low vs. high) to evaluate the predictive performance of the radiomics model in different subgroups. This study was approved by the Ethics Committee of the First Affiliated Hospital of Soochow University (Approval No. 2011197), and written informed consent was obtained from all participants. The detailed research methodology is depicted in Figure 2.
Figure 2.
Research workflow diagram.
CDU protocol
A standardized ultrasound protocol with preset parameters was employed to ensure data accuracy and consistency. CDU examinations were performed using the iU Elite scanner (Philips Healthcare, Netherlands), with L9-3 linear array and C5-1 convex array probes scanning from proximal to distal vascular segments. Transverse and longitudinal views were acquired using grayscale imaging, color flow imaging, and spectral Doppler analysis to assess the bilateral common carotid arteries, carotid artery bifurcation (CAB), and internal carotid arteries (ICA). When plaques were present, their dimensions, morphology, echogenicity, structural characteristics, and degree of vascular stenosis were evaluated from multiple angles and planes. Key hemodynamic parameters were recorded, such as peak systolic velocity, end-diastolic velocity, and resistive index. Images of carotid atherosclerosis (CAS) related plaques were archived for subsequent analysis.
Analysis of CDU
In this study, the criteria for plaque characterization were based on our team’s previous pathological research. Radiomic features were extracted from CDU images and integrated with clinical data to develop a predictive model. The workflow included feature quality control using the intraclass correlation coefficient (ICC), feature selection via least absolute shrinkage and selection operator (LASSO) regression, and model construction using machine learning algorithms (Figure 3).
Figure 3.
Radiomics feature extraction flowchart.
Plaques were classified according to a vulnerability score, with a score of 4 indicating vulnerable plaques and scores other than 4 classified as stable. The scoring system was based on ultrasound features such as hypoechogenicity and intraplaque hemorrhage, combined with histological markers like lipid-rich necrotic core and hemorrhage as defined in the literature (26). Two-dimensional ultrasound was used to assess plaque morphology, echogenicity, fibrous cap integrity, and the presence of ulceration (e.g., surface depression >1 mm or visible intraplaque color flow), as well as the degree of stenosis, which was graded from 0 to 3 (27). Due to the limited resolution of CDU in distinguishing lipid core from intraplaque hemorrhage, these features were analyzed as a combined variable (22).
Subsequently, the most prognostically relevant radiomic features were selected to enhance the robustness and clinical applicability of the model. The detailed process is illustrated in the diagram (Figure 4), which outlines the key steps from feature selection to model development, emphasizing the role of statistical and machine learning techniques in processing high-dimensional imaging data to construct a highly accurate and generalizable radiomic nomogram.
Figure 4.
Processing and model building flowchart. This flowchart illustrates the stepwise procedure for selecting key radiomic features from the CDU data and integrating them with clinical factors to construct the predictive model. It encompasses the use of ICC for feature quality control, LASSO regression for feature selection, and the application of machine learning algorithms to establish a prognostic nomogram.
Image segmentation and radiomic feature calculation
The maximum longitudinal section of each carotid plaque was identified during an ultrasound examination based on size, shape, and spatial relationship to surrounding structures, using the iU Elite scanner (Philips Healthcare, Netherlands) for comprehensive multi-angle CDU (carotid duplex ultrasound) scanning. Regions of interest (ROIs) were manually delineated on the maximum longitudinal section using ITK-SNAP 4.0.0-BETA by two experienced radiologists independently and in a blinded manner (Figure 5). Inter-observer agreement was excellent (Cohen’s kappa = 0.93). To minimize variability and ensure feature robustness, a three-step normalization strategy was applied: spatial normalization (resampling to 0.2 × 0.2 mm2 resolution), intensity normalization using histogram matching, and morphological compensation by incorporating plaque metrics (length, area, perimeter ratio) as covariates in LASSO regression.
Figure 5.
Example of manually segmented CDU images. (A) Original ultrasound image of the plaque. (B) The segmentation region was adjusted along the maximum longitudinal cross-section of the plaque. (C) ITK-SNAP software was used to label the segmented plaque image for further analysis.
Radiomic features were extracted from both the original CDU images and a series of filtered images. The following image filters were applied: wavelet transform (eight decompositions), Laplacian of Gaussian (LoG) with sigma values of 1.0, 2.0, 3.0, 4.0, and 5.0, Local Binary Patterns (LBP) in 3D, square filter, square root filter, logarithm filter, exponential filter, and gradient filter. From the standardized ROIs (original and filtered images), a total of 1,477 radiomic features were extracted, covering shape, intensity, and texture characteristics. Texture features were derived using methods including gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and gray level dependence matrix (GLDM) (28). Highly collinear features (Pearson correlation coefficient > 0.9) were excluded to reduce redundancy. Feature definitions and extraction protocols are available at https://pyradiomics.readthedocs.io/.
Feature reduction and model development
The entire dataset was first randomly split into a training set (70%, n = 74) and a testing set (30%, n = 31). All feature reduction and model development procedures were performed exclusively on the training set; the testing set was kept completely independent and was not used until the final model evaluation.
In the training set, radiomic features were preliminarily screened. Features with low reproducibility (intraclass correlation coefficient, ICC < 0.75) and high collinearity (Pearson correlation coefficient > 0.90) were excluded to retain informative and independent variables. Spearman analysis showed only 2 of 1,477 geometric features (equivalent diameter, maximum 3D diameter) were strongly correlated with plaque size (ρ > 0.6) and were removed during preprocessing. The remaining features were selected using LASSO regression, with 10-fold cross-validation to determine the optimal penalty. During LASSO cross-validation, feature selection was repeated within each training fold to prevent information leakage from the validation folds. The final model incorporated CDU radiomic features, systemic immune-inflammation index (SII), and clinical variables such as NIHSS score.
Model performance was evaluated using logistic regression (LR), k-nearest neighbor (KNN), and Light Gradient Boosting Machine (LightGBM). LR provides strong interpretability and is suitable for moderate-sized datasets but assumes linear relationships. KNN handles nonlinear data well but is less efficient with large datasets and is sensitive to noise. LightGBM efficiently manages complex, high-dimensional data with automatic feature selection but requires careful tuning to prevent overfitting. The model choice was based on data characteristics and clinical relevance. Performance was assessed via k-fold cross-validation. The dataset was split into k subsets, with each used once as a test set while the others served for training. Metrics such as accuracy and recall were averaged to evaluate model consistency. Discriminative ability was quantified by the area under the ROC curve (AUC), which is robust to class imbalance and widely used in clinical prediction models.
Validation of the radiomics column chart model
After the feature selection and model building were finalized in the training set, the testing set was used solely for model evaluation. The model was rigorously validated in training and testing sets using ROC, calibration curves, decision curve analysis (DCA), and Brier scores to evaluate discrimination, calibration, and clinical utility. Calibration curves assessed the agreement between predicted and observed outcomes, while Brier scores quantified predictive accuracy by measuring the mean squared difference between predicted probabilities and actual outcomes. Lower Brier scores indicated better calibration, particularly in the testing set, supporting the model’s reliability in clinical applications.
Statistical analysis
Statistical analysis was performed using SPSS version 26.0 and Python 3.6.7. The dataset was randomly divided into a training set (70%) and a testing set (30%) before any feature selection or model development. All feature reduction steps (ICC, Pearson, LASSO) were performed exclusively on the training set. Normally distributed data were expressed as mean ± standard deviation, and non-normally distributed data as median with interquartile range. Categorical variables were presented as percentages. Group comparisons were conducted using chi-square or Fisher’s exact tests for categorical data, and t-tests or Mann–Whitney U tests for continuous variables, as appropriate. Univariate and multivariate logistic regression analyses were used to identify independent clinical predictors and to develop the predictive model. Model performance was assessed using ROC curves, calibration curves, and DCA. Inter-rater agreement for image segmentation was evaluated using Cohen’s kappa coefficient. All statistical tests were two-sided, with significance set at p = 0.05.
Results
Patient characteristics
Participant eligibility was determined based on predefined inclusion and exclusion criteria. One hundred five AIS patients were included, with a mean age of 62.45 ± 10.38 years; 72 were male (68.57%). Among them, 60 patients (57.14%) were diagnosed with hypertension and 47 (44.76%) with diabetes. Baseline characteristics were categorized according to clinical outcomes (favorable vs. unfavorable), as shown in Table 1. Significant differences were observed between the two groups in age, diabetes, NIHSS score, HCY levels, and presence of vulnerable plaques (p = 0.05). After excluding patients with atrial fibrillation, radiomic features of carotid plaques remained significantly associated with stroke outcomes (p = 0.05). In the anterior circulation subgroup, radiomic features were significantly associated with unfavorable outcomes (p = 0.05), while no significant association was observed in the posterior circulation subgroup (p = 0.12).
Table 1.
Baseline data of patients according to poor/good prognosis (N = 105).
| Characteristics | Poor prognosis (n = 78) | Good prognosis (n = 27) | p-value |
|---|---|---|---|
| Age (median ± SD) | 64.23 ± 10.76 | 62.89 ± 10.14 | 0.034* |
| Male (%) | 52/26 | 20/7 | 0.456 |
| Hypertension (%) | 48/30 | 12/15 | 0.563 |
| Systolic blood pressure (median ± SD) | 136.48 ± 17.12 | 129.14 ± 14.28 | 0.065 |
| Diastolic blood pressure (median ± SD) | 83.44 ± 10.91 | 81.73 ± 10.12 | 0.234 |
| Diabetes mellitus (%) | 36/42 | 11/16 | 0.045* |
| Coronary heart disease (%) | 38/40 | 13/14 | 0.865 |
| Current or former smokers (%) | 25/53 | 12/15 | 0.340 |
| History of alcohol intake (%) | 15/63 | 5/22 | 0.567 |
| History statin treatment (%) | 24/54 | 8/19 | 0.432 |
| History antiplatelet treatment (%) | 30/48 | 9/18 | 0.125 |
| History of stroke (%) | 27/51 | 7/20 | 0.456 |
| NIHSS score [Median (Q1 – Q3)] | 8.00 [6.00,13.00] | 6.40 [4.00,8.50] | 0.001* |
| Fasting blood-glucose (median ± SD) | 5.91 [5.00,7.28] | 5.88 [5.45,7.15] | 0.678 |
| Dyslipidemia (%) | 28/50 | 8/19 | 0.021* |
| TC (median ± SD) | 4.07 ± 1.04 | 4.39 ± 1.23 | 0.567 |
| TG (median ± SD) | 1.26 [0.90,1.66] | 1.34 [0.92,1.72] | 0.530 |
| HDL-C (median ± SD) | 1.07 [0.91,1.29] | 1.15 [0.97,1.31] | 0.165 |
| LDL-C [Median (Q1 – Q3)] | 2.52 [1.85,3.15] | 2.72 [1.95,3.17] | 0.208 |
| Uric acid (median ± SD) | 320.25 ± 84.55 | 310.18 ± 94.33 | 0.345 |
| Fibrinogen [Median (Q1 – Q3)] | 3.21 [2.62,3.79] | 3.15 [2.63,4.01] | 0.876 |
| HCY (median ± SD) | 9.42 [7.48,12.84] | 7.95 [5.42,10.44] | 0.034* |
| hs-CRP [Median (Q1 – Q3)] | 2.69 [1.72,5.32] | 2.70 [1.50,4.85] | 0.675 |
| Platelet counts [Median (Q1 – Q3)] | 210.50 [172.00,244.75] | 193.00 [148.50,235.00] | 0.076 |
| Neutrophil counts [Median (Q1 – Q3)] | 4.73 [3.65,6.03] | 4.39 [3.45,5.36] | 0.078 |
| Lymphocyte counts [Median (Q1 – Q3)] | 1.68 [1.35,2.15] | 1.75 [1.34,2.30] | 0.453 |
| Monocyte counts [Median (Q1 – Q3)] | 0.53 [0.37,0.62] | 0.50 [0.33,0.57] | 0.298 |
| SII [Median (Q1 – Q3)] | 589.70 [400.82,832.27] | 478.50 [295.21,679.79] | 0.013* |
| PLR [Median (Q1 – Q3)] | 123.47[96.60,158.82] | 114.84 [79.76,138.80] | 0.056 |
| MLR [Median (Q1 – Q3)] | 0.28 [0.19,0.40] | 0.26 [0.18,0.37] | 0.132 |
| NLR [Median (Q1 – Q3)] | 2.70 [1.96,3.96] | 2.35 [1.68,3.44] | 0.102 |
| Vulnerable Plaque (%) | 52/26 | 12/15 | <0.001* |
| Ulcerative plaque | 50/28 | 9/18 | <0.001* |
| Hypoechoic plaque | 37/41 | 8/19 | 0.897 |
| Irregular plaque | 49/29 | 10/17 | <0.001* |
| Infarct volume (cm3) | 18.3 ± 12.1 | 32.7 ± 20.5 | <0.001* |
| ASPECTS scoring | 8.2 ± 1.5 | 5.9 ± 2.3 | 0.003 |
* indicates p-values indicating statistical significance. TC, total cholesterol; TG, triglycerides; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; NIHSS, National Institutes of Health Stroke Scale; HCY, homocysteine; hs-CRP, high-sensitive C-reactive protein; SII, systemic immune-inflammation index; PLR, platelet to lymphocyte ratio; MLR, monocyte to lymphocyte ratio; NLR, neutrophil to lymphocyte ratio.
Comparison of clinical and imaging characteristics between poor prognosis and good prognosis groups
Early identification of prognostic factors is critical in the management of AIS. This study analyzed clinical and biomarker parameters associated with favorable and unfavorable outcomes. Table 1 summarizes the baseline characteristics of 105 AIS patients grouped by outcome. Significant differences were observed in age, diabetes, dyslipidemia, NIHSS score, HCY, SII, and plaque features, including vulnerable, ulcerative, and irregular plaques (p = 0.05). As shown in Figure 6, the mean age was higher in the unfavorable group (64.23 years, n = 78) than in the favorable group (62.48 years, n = 27) (p = 0.044). The unfavorable group had a higher prevalence of diabetes (p = 0.049) and elevated HCY levels (p = 0.042). The median NIHSS score was also higher (8 vs. 6.4, p = 0.001), along with SII (p = 0.015). Vulnerable, ulcerative, and irregular plaques were significantly more common in the unfavorable group (p = 0.001). Subgroup analysis indicated a significant correlation between the radiomic feature GLCM_Contrast and infarct volume in the anterior circulation group (r = 0.62, p = 0.001), but not in the posterior group (r = 0.11, p = 0.42). The anterior circulation-specific nomogram achieved an AUC of 0.902 (95% CI: 0.87–0.95), significantly outperforming the unstratified model (ΔAUC = 0.036, p = 0.021) (Figure 7).
Figure 6.
Comparison of clinical and radiological features between poor prognosis group and good prognosis group. The blue bars in the figure represent the good prognosis group, while the orange bars represent the poor prognosis group.
Figure 7.
Correlation heatmaps of radiomic features and ROC curve comparisons between anterior and posterior circulation stroke models. (A) The heatmap illustrates significant intergroup differences in intraplaque hemorrhage (IPH), fibrous cap thickness, peak systolic velocity (PSV), and end-diastolic velocity (EDV) (*p = 0.05, **p = 0.01). (B) ROC curves compare the predictive performance of the overall model, anterior circulation model, and posterior circulation model.
Feature selection and radiomics signature development
From each plaque, 1,477 ultrasound radiomics features were initially extracted. Following inter-observer consistency evaluation, 88 features with an ICC < 0.75 were excluded. Pearson correlation analysis was then performed to assess feature collinearity, and 1,016 features with Pearson correlation coefficients < 0.9 were retained (Figure 8). Utilizing the intercept and coefficients derived from LR, we established the radiomics scoring formula as follows: Radscore = 0.7883–0.0136exponential_firstorder_InterquartileRange-0.0030exponential_glcm_InverseVariance+0.0303exponential_glrlm_ShortRunHighGrayLevelEmphasis-0.008exponential_glrlm_ShortRunLowGrayLevelEmphasis-0.0219exponential_glszm_GrayLevelNonUniformity-0.0109gradient_glrlm_LongRunHighGrayLevelEmphasis-0.0161lbp_3D_m1_gldm_DependenceVariance-0.0251lbp_3D_m2_firstorder_InterquartileRange-0.0015square_glrlm_ShortRunLowGrayLevelEmphasis-0.0062squareroot_gldm_DependenceVariance+0.0284*wavelet_HLL_glcm_Correlation.
Figure 8.
Selection process of radiomics features. (A) Variation of variable coefficients with penalty coefficients. (B) The optimal penalty coefficient is selected through 10-fold cross-validation. When the binomial deviation is minimized (the minimum standard), 11 optimal radiomics features are chosen. (C) Absolute values of the included variable coefficients.
Multifactorial predictive modeling analysis of adverse outcomes in AIS patients
Univariate and multivariate logistic regression analyses investigated various clinical indicators’ independent contributions and interrelationships in prognostic assessment. As presented in Table 2, univariate analysis revealed that age (OR = 1.284, 95% CI: 1.005–1.399, p = 0.045), diabetes mellitus (OR = 1.722, 95% CI: 1.003–2.091, p = 0.049), and dyslipidemia (OR = 1.812, 95% CI: 1.019–2.027, p = 0.031) were significantly associated with poor prognosis. In multivariate analysis, age remained statistically significant with a slightly lower OR (OR = 1.280, 95% CI: 1.002–1.391, p = 0.046).
Table 2.
Univariate and multivariate logistic regression analysis of factors related to poor prognosis (N = 105).
| Characteristics | Univariate | Multivariate | ||
|---|---|---|---|---|
| OR [95% CI] | p | OR [95% CI] | p | |
| Age | 1.284 [1.005–1.399] | 0.045* | 1.280 [1.002–1.391] | 0.046* |
| Systolic blood pressure | 1.015 [0.993–1.036] | 0.078 | – | – |
| Diabetes mellitus | 1.722 [1.003–2.091] | 0.049* | – | – |
| Dyslipidemia | 1.812 [1.019–2.027] | 0.031* | – | – |
| NHISS score | 6.102 [2.005–18.800] | 0.003* | 2.532 [1.901–3.102] | <0.001* |
| Platelet counts | 1.002 [0.999–1.006] | 0.088 | – | – |
| Neutrophil counts | 1.230 [0.980–1.523] | 0.060 | – | – |
| SII | 2.952 [2.105–3.620] | 0.020* | 2.955 [2.251–3.509] | 0.017* |
| PLR | 1.006 [1.000–1.013] | 0.051* | – | – |
| Vulnerable Plaque | 1.220 [1.110–1.387] | <0.001* | 1.225 [1.098–1.379] | <0.001* |
| Ulcerative plaque | 2.174 [2.070–2.420] | <0.001* | – | – |
| Irregular plaque | 1.258 [1.121–1.541] | <0.001* | – | – |
* indicates P-values indicating statistical significance. OR, Odds ratio; CI, Confidence interval; NIHSS, National Institutes of Health Stroke scale; HCY, Homocysteine; hs-CRP, high-sensitive C-reactive protein; SII, systemic immune-inflammation index.
Furthermore, the NIHSS score and the SII emerged as independent predictors of adverse outcomes in univariate and multivariate analyses. The univariate OR for the NIHSS score was 6.102 (95% CI: 2.005–18.800, p = 0.003), and the multivariate OR was 2.532 (95% CI: 1.901–3.102, p < 0.001). For the SII, the univariate OR was 2.952 (95% CI: 2.105–3.620, p = 0.020), and the multivariate OR was 2.955 (95% CI: 2.251–3.509, p = 0.017). Vulnerable plaque also demonstrated significant association with poor prognosis in both analyses, with a univariate OR of 1.220 (95% CI: 1.110–1.387, p < 0.001) and a multivariate OR of 1.225 (95% CI: 1.098–1.379, p < 0.001).
Construction of ultrasound radiomics column chart
A radiomics nomogram was developed by integrating four clinical features with 11 ultrasound radiomic features (Figure 9). In the testing set, the model achieved a Brier score of 0.104, indicating a small deviation between predicted probabilities and actual outcomes. ROC curves comparing the clinical model, radiomic model, and radiomics nomogram. AUC values for three classifiers (LightGBM, LR, and KNN) based on the radiomic feature model in both training and testing sets. AUC values for the clinical model constructed with four clinical variables are provided in Supplementary Figure S1.
Figure 9.
Ultrasonic radiomics column chart. In the training set, a chart encompassing both imaging features and clinical characteristics was constructed. The sum of scores for each variable yields a total score corresponding to the probability of predicting adverse outcomes. This allows for the direct assessment of whether each patient in the chart is at risk for adverse outcomes.
The radiomics nomogram yielded an AUC of 0.891 (95% CI: 0.832–0.950) in the training set and 0.852 (95% CI: 0.757–0.947) in the testing set. In the training set, the AUCs for the clinical model, radiomic model, and nomogram were 0.794, 0.879, and 0.891, respectively; in the testing set, the corresponding values were 0.765, 0.864, and 0.852 (Table 3). Accuracy, sensitivity, and specificity metrics demonstrated superior predictive performance for the radiomics nomogram. DeLong test results indicated statistically significant differences in AUCs between the nomogram and the other models in both sets (p = 0.05). The AUC values of the LR model were consistent across training and testing sets.
Table 3.
Prediction efficiency indexes of the clinical signature, radiomic signature, and radiomic nomogram in the train and test set.
| Signature | AUC | 95% CI | Accuracy | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|---|---|
| Training group | 0.794 | 0.707–0.881 | 0.781 | 0.728 | 0.642 | 0.703 | 0.648 |
| Clinic Signature | 0.879 | 0.817–0.941 | 0.825 | 0.723 | 0.790 | 0.796 | 0.726 |
| Rad Signature | 0.891 | 0.832–0.950 | 0.829 | 0.756 | 0.799 | 0.884 | 0.758 |
| Testing group | |||||||
| Clinic Signature | 0.765 | 0.577–0.953 | 0.689 | 0.701 | 0.675 | 0.690 | 0.614 |
| Rad Signature | 0.864 | 0.766–0.962 | 0.774 | 0.785 | 0.818 | 0.798 | 0.736 |
| Nomogram | 0.852 | 0.757–0.947 | 0.789 | 0.805 | 0.777 | 0.889 | 0.762 |
AUC, area under curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value. After propensity score matching (n = 86), the radiomics signature maintained comparable predictive performance with AUC 0.861 (95%CI: 0.785–0.937). No significant differences remained in age (p = 0.32), NIHSS (p = 0.26), SII (p = 0.41) or vulnerable plaque prevalence (p = 0.19) between matched groups.
The application value of clinical physicians
A radionics-based prediction model was developed using carotid ultrasound images in AIS patients, and its performance was evaluated using multiple statistical methods. As shown in Figure 10, the calibration curve demonstrated good agreement between predicted probabilities and actual outcomes for adverse prognosis within 3 months. Model performance was consistent across training and testing datasets, with the Hosmer–Lemeshow test supporting its calibration. DCA in Supplementary Figure S2 indicated that the radiomics model yielded a higher net benefit than traditional clinical models across a range of threshold probabilities. Inter-observer agreement for plaque vulnerability assessment using CDU was high, with a Cohen’s kappa of 0.93 (95% CI: 0.87–0.97). Agreement for ulcerative plaques was 0.90 (95% CI: 0.81–0.97), for lipid necrotic core/intraplaque hemorrhage was 0.88 (95% CI: 0.80–0.96), and for fibrous cap rupture/incompleteness was 0.83 (95% CI: 0.74–0.94).
Figure 10.
Calibration curves of different models. This figure displays the calibration curves of various models in both the training and testing datasets, showcasing the consistency between the predicted probabilities of adverse outcomes by the models and the actual observed probabilities. The calibration curves demonstrate a strong alignment between the predicted risks of adverse outcomes in stroke patients within 3 months and the observed risks in both the training and testing datasets.
Discussion
This study is the first to integrate CDU radiomic features with the SII and key clinical variables to construct a multimodal prediction model for assessing three-month outcomes in patients with AIS. Compared to high-resolution MRI, which has advantages in plaque composition analysis but is costly and time-consuming (28), our approach utilizes low-cost (approximately 50 RMB) and rapid (<10 min) CDU imaging to extract quantifiable high-dimensional radiomic features, significantly enhancing the identification of plaque heterogeneity and vulnerability (AUC = 0.876). This model offers strong predictive performance, practical utility, and high scalability, particularly for early risk stratification and precision management in low-resource settings.
The stratified analysis confirmed that age, NIHSS score, SII, and vulnerable plaques are independent predictors of poor prognosis at 3 months. Strict inclusion criteria were applied to improve data quality and model accuracy, limiting the cohort to patients admitted within 3 days of onset without severe comorbidities or atrial fibrillation. After propensity score matching, the model maintained strong predictive ability (AUC = 0.861), supporting its independence and complementary value to clinical variables. The model performed better in anterior circulation strokes, consistent with the anatomical role of carotid plaques in supplying the anterior territory, while the posterior circulation may be more affected by systemic inflammation. To further explain this difference, several physiological factors should be considered. The posterior circulation (vertebrobasilar system) is primarily supplied by the vertebral arteries rather than the carotid arteries. Embolic events resulting from carotid plaque rupture or shedding are more likely to affect the anterior circulation (middle cerebral artery and anterior cerebral artery territories), as the internal carotid artery directly supplies these regions. In contrast, posterior circulation ischemia involves more complex and heterogeneous pathological mechanisms, including perforator artery disease, vertebrobasilar dolichoectasia, and anomalies in vertebral artery origin, which are less directly related to carotid plaque characteristics. These factors may dilute the predictive value of carotid ultrasound radiomic features for posterior circulation strokes. Additionally, the relatively small number of posterior circulation cases in our cohort may have limited the statistical power to detect significant associations in this subgroup. This mechanistic heterogeneity reinforces the biological plausibility of the model. Previous studies have shown higher mortality and poorer recovery in elderly AIS patients, likely due to age-related decline and complications such as infections or hemorrhagic transformation (29, 30). The prognostic relevance of NIHSS and the role of inflammation in AIS progression have also been established (31). By excluding patients with atrial fibrillation, our study confirms that carotid plaque imaging features remain significant prognostic markers in non-cardioembolic stroke.
The SII, which integrates neutrophil, lymphocyte, and platelet counts, has been widely used to predict outcomes in various diseases (32). In AIS, neutrophils exacerbate brain injury by releasing pro-inflammatory mediators such as matrix metalloproteinases (33, 34), while regulatory T cells and certain T-cell subsets may exert neuroprotective effects through anti-inflammatory pathways. Platelet activation contributes to thrombosis and inflammation associated with cerebral ischemia (35, 36). In this study, SII demonstrated independent prognostic value as a global marker of inflammation.
Vulnerable carotid plaques also play a critical role in AIS progression, with previous studies indicating associations with prolonged hospitalization and poor functional recovery (37, 38). To eliminate confounding from baseline disability, patients with a modified Rankin Scale score greater than 2 were excluded, ensuring a more accurate assessment of plaque characteristics. Radiomics enabled the quantification of texture, grayscale, and morphological heterogeneity of plaques beyond visual assessment (38–40). Features like GLCM and GLRLM captured microstructural complexity and demonstrated a low correlation with subjective vulnerability scores, supporting their independent predictive value. The ROI was standardized to the largest longitudinal section of the plaque, with blind segmentation and inter-observer agreement testing to ensure feature stability (9). Mechanistically, metabolic disturbances and inflammation contribute to the formation of vulnerable plaques, which elevate the risk of thrombosis and cerebral ischemia, ultimately leading to poor AIS outcomes (41, 42). While several AIS prognostic models have been developed (43–45), many suffer from limitations in scoring systems or inconsistent variable definitions. By integrating radiomics with clinical and immunological markers, this study proposes a prognostic tool with enhanced biological plausibility and clinical applicability.
Previous studies have developed prognostic scoring systems for AIS using multivariate logistic regression. For example, the Acute Stroke Registry and Lausanne Score identified age, NIHSS score, time from onset to admission, visual field deficits, blood glucose, and level of consciousness as key variables (45). However, some of these predictors rely on subjective assessment or are unsuitable for patients with impaired consciousness, limiting their generalizability. In contrast, our model incorporates the modified Rankin Scale (mRS), NIHSS, SII and radiomic features derived from CDU, offering a more clinically adaptable tool. NIHSS reflects the severity of neurological deficits, and SII captures the systemic inflammatory burden. Together with plaque imaging features such as ulceration or irregular borders, these variables provide a multidimensional framework for outcome prediction. Integrating radiomic nomograms and conventional clinical indicators demonstrated superior performance and practical potential compared to traditional scoring systems (Figure 11).
Figure 11.
Analysis of radiomics and clinical factors in prognosis of AIS patients.
Regarding model development, LASSO regression was used for feature selection, followed by model construction and comparison using three machine learning algorithms: LightGBM, LR, and KNN. LR achieved the best performance in the radiomics model, combining high training efficiency with interpretability and the ability to quantify the impact of predictors through odds ratios, making it highly suitable for clinical application (46). Although KNN is advantageous for nonlinear data, it is highly sensitive to class imbalance, which likely contributed to its suboptimal performance in this study. Overall, the CDU-based radiomics model outperformed other methods in training and testing cohorts and demonstrated good calibration consistency, supporting its reliability and robustness in AIS prognostic assessment (47).
Despite the innovative nature and clinical applicability of this study, several limitations should be acknowledged. First, as a prospective cohort with retrospective radiomic analysis, there is a potential risk of selection bias. Second, the relatively small sample size may limit the generalizability of the model. Incomplete acquisition of key imaging data, such as infarct location, volume, and collateral circulation status, may affect predictive performance across varying lesion characteristics. The retrospective nature of the radiomic analysis (i.e., feature extraction and model development performed on prospectively collected data) also limited the evaluation of dynamic infarct evolution, such as 24-h volume expansion, which future studies should address using longitudinal radiomic data.
Plaque vulnerability scores in this study were inferred primarily from imaging features. Although double-blind segmentation and inter-observer agreement assessments were conducted to reduce subjectivity, further histopathological validation is warranted to enhance accuracy. Additionally, external validation using independent cohorts is lacking.
Furthermore, although our subgroup analysis showed better performance in anterior circulation strokes, the inclusion of both anterior and posterior circulation cases in the primary model may have introduced some heterogeneity.
Future research should include multicenter, prospective studies to systematically assess model performance across diverse populations and imaging platforms. Expanding the sample size and refining feature selection and modeling algorithms will improve prediction accuracy and facilitate broader clinical implementation of AIS prognostic tools.
Acknowledgments
The machine learning algorithms (LASSO, LightGBM, KNN, and LR) were used solely for predictive model construction and performance evaluation.
Funding Statement
The author(s) declared that financial support was received for this work and/or its publication. This study was supported by the Key Scientific Research Project of Yancheng Municipal Health Commission (YK2024033).
Footnotes
Edited by: Jean-Claude Baron, University of Cambridge, United Kingdom
Reviewed by: Caleb E. Feliciano, University of Puerto Rico, Medical Sciences Campus, Puerto Rico
Ahmadreza Sohrabi-Ashlaghi, Tehran University of Medical Sciences, Iran
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Ethics statement
The studies involving humans were approved by the Ethics Committee of the First Affiliated Hospital of Soochow University (Approval number: 2011197). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.
Author contributions
LZ: Funding acquisition, Investigation, Methodology, Writing – original draft, Writing – review & editing. HC: Data curation, Methodology, Software, Writing – review & editing. LM: Data curation, Methodology, Software, Validation, Writing – review & editing. XZ: Investigation, Methodology, Project administration, Supervision, Writing – review & editing. GS: Conceptualization, Methodology, Project administration, Supervision, Writing – review & editing.
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/fneur.2026.1878428/full#supplementary-material
ROC curves of the clinical model in the training and testing sets. Area under the curve (AUC).
DCA of clinical features, radiomic features, and radiomic nomogram. This figure illustrates the DCA results of clinical features, radiomic features, and radiomic nomograms, comparing the efficacy of these models in clinical decision-making. In comparison to scenarios without using any predictive models (all-treatment or no-treatment strategies), a radiomic nomogram demonstrates significant intervention benefits in patients with predictive probabilities, as opposed to clinical features and radiomic features. DCA stands for decision curve analysis.
References
- 1.Zhang J, Zhao J. Clinical characteristics and analysis of lung cancer-associated acute ischemic stroke. J Stroke Cerebrovasc Dis. (2020) 29:105164. doi: 10.1016/j.jstrokecerebrovasdis.2020.105164, [DOI] [PubMed] [Google Scholar]
- 2.Feigin VL, Stark BA, Johnson CO, Roth GA, Bisignano C, Abady GG, et al. Global, regional, and national burden of stroke and its risk factors, 1990–2019: a systematic analysis for the global burden of disease study 2019. Lancet Neurol. (2021) 20:795–820. doi: 10.1016/s1474-4422(21)00252-0, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Tu WJ, Wang LD, Yan F, Peng B, Hua Y, Liu M, et al. China stroke surveillance report 2021. Mil Med Res. (2023) 10:33. doi: 10.1186/s40779-023-00463-x, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Ding Q, Liu S, Yao Y, Liu H, Cai T, Han L. Global, regional, and National Burden of ischemic stroke, 1990–2019. Neurology. (2022) 98:e279–90. doi: 10.1212/wnl.0000000000013115, [DOI] [PubMed] [Google Scholar]
- 5.Chaudhary D, Abedi V, Li J, Schirmer CM, Griessenauer CJ, Zand R. Clinical risk score for predicting recurrence following a cerebral ischemic event. Front Neurol. (2019) 10:10. doi: 10.3389/fneur.2019.01106, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Roth GA, Abate D, Abate KH, Abay SM, Abbafati C, Abbasi N, et al. Global, regional, and national age-sex-specific mortality for 282 causes of death in 195 countries and territories, 1980–2017: a systematic analysis for the global burden of disease study 2017. Lancet. (2018) 392:1736–88. doi: 10.1016/s0140-6736(18)32203-7, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Elmusrati M, Ashammakhi N. Cancer-on-a-Chip and Artificial intelligence: tomorrow’s Cancer management. J Craniofacial Surg. (2018) 29:1682–3. doi: 10.1097/scs.0000000000004703, [DOI] [PubMed] [Google Scholar]
- 8.Ge Y, Wang Q, Wang L, Wu H, Peng C, Wang J, et al. Predicting post-stroke pneumonia using deep neural network approaches. Int J Med Inform. (2019) 132:103986. doi: 10.1016/j.ijmedinf.2019.103986, [DOI] [PubMed] [Google Scholar]
- 9.Zhang L, Lyu Q, Ding Y, Hu C, Hui P. Texture analysis based on vascular ultrasound to identify the vulnerable carotid plaques. Front Neurosci. (2022) 16:885209. doi: 10.3389/fnins.2022.885209, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Yan Y, Lu Z, Ding Y, Pu J, Hu C, Teng Z, et al. Predictors of ischemic events in patients with unilateral extracranial vertebral artery dissection: a single-center exploratory study. Front Neurol. (2022) 13:939001. doi: 10.3389/fneur.2022.939001, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Ye Z, Hu T, Wang J, Xiao R, Liao X, Liu M, et al. Systemic immune-inflammation index as a potential biomarker of cardiovascular diseases: A systematic review and meta-analysis. Front Cardiovasc Med. (2022) 9:9. doi: 10.3389/fcvm.2022.933913, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Bao Y, Wang L, Du C, Ji Y, Dai Y, Jiang W. Association between systemic immune inflammation index and cognitive impairment after acute ischemic stroke. Brain Sci. (2023) 13:464. doi: 10.3390/brainsci13030464, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Cui Z, Kuang S, Yang X, Wang Y, Gu S, Li H, et al. Predictive value of the systemic immune inflammation (SII) index for stroke-associated pneumonia. Brain Behav. (2023) 13:e3302. doi: 10.1002/brb3.3302, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Famakin B. The immune response to acute focal cerebral ischemia and associated post-stroke immunodepression: a focused review. Aging Dis. (2014) 5:307–26. doi: 10.14336/ad.2014.0500307 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Zhang D, Ren J, Luo Y, He Q, Zhao R, Chang J, et al. T cell response in ischemic stroke: from mechanisms to translational insights. Front Immunol. (2021) 12:12. doi: 10.3389/fimmu.2021.707972, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Fann DYW, Lee SY, Manzanero S, Chunduri P, Sobey CG, Arumugam TV. Pathogenesis of acute stroke and the role of inflammasomes. Ageing Res Rev. (2013) 12:941–66. doi: 10.1016/j.arr.2013.09.004, [DOI] [PubMed] [Google Scholar]
- 17.Wang Z, Chen G, Chen Z q, Mou R t, Feng D x. The role of nitric oxide in stroke. Med Gas Res. (2017) 7:194. doi: 10.4103/2045-9912.215750, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Yamada K, Kawasaki M, Yoshimura S, Shirakawa M, Uchida K, Shindo S, et al. High-intensity signal in carotid plaque on routine 3D-TOF-MRA is a risk factor of ischemic stroke. Cerebrovasc Dis. (2015) 41:13–8. doi: 10.1159/000441094, [DOI] [PubMed] [Google Scholar]
- 19.Brinjikji W, Huston J, Rabinstein AA, Kim GM, Lerman A, Lanzino G. Contemporary carotid imaging: from degree of stenosis to plaque vulnerability. JNS. (2016) 124:27–42. doi: 10.3171/2015.1.jns142452, [DOI] [PubMed] [Google Scholar]
- 20.Tagliafico AS, Piana M, Schenone D, Lai R, Massone AM, Houssami N. Overview of radiomics in breast cancer diagnosis and prognostication. Breast. (2020) 49:74–80. doi: 10.1016/j.breast.2019.10.018, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.van Engelen A, Wannarong T, Parraga G, Niessen WJ, Fenster A, Spence JD, et al. Three-dimensional carotid ultrasound plaque texture predicts vascular events. Stroke. (2014) 45:2695–701. doi: 10.1161/strokeaha.114.005752, [DOI] [PubMed] [Google Scholar]
- 22.van Veelen A, van der Sangen NMR, Delewi R, Beijk MAM, Henriques JPS, Claessen BEPM. Detection of vulnerable coronary plaques using invasive and non-invasive imaging modalities. JCM. (2022) 11:1361. doi: 10.3390/jcm11051361, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Devi M, Singh S, Tiwari S, Chandra Patel S, Ayana MT. A survey of soft computing approaches in biomedical imaging. J Healthcare Engineer. (2021) 2021:1–15. doi: 10.1155/2021/1563844, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Zhang Y, Wang S. An innovative forecasting model to predict wind energy. Environ Sci Pollut Res. (2022) 29:74602–18. doi: 10.1007/s11356-022-20971-y, [DOI] [PubMed] [Google Scholar]
- 25.Davies RH, Augusto JB, Bhuva A, Xue H, Treibel TA, Ye Y, et al. Precision measurement of cardiac structure and function in cardiovascular magnetic resonance using machine learning. J Cardiovasc Magn Reson. (2022) 24:16. doi: 10.1186/s12968-022-00846-4, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Saam T, Habs M, Buchholz M, Schindler A, Bayer-Karpinska A, Cyran CC, et al. Expansive arterial remodeling of the carotid arteries and its effect on atherosclerotic plaque composition and vulnerability: an in-vivo black-blood 3T CMR study in symptomatic stroke patients. J Cardiovasc Magn Reson. (2016) 18:11. doi: 10.1186/s12968-016-0229-2, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Zhou R, Luo Y, Fenster A, Spence JD, Ding M. Fractal dimension based carotid plaque characterization from three-dimensional ultrasound images. Med Biol Eng Comput. (2018) 57:135–46. doi: 10.1007/s11517-018-1865-5, [DOI] [PubMed] [Google Scholar]
- 28.Varghese BA, Fields BKK, Hwang DH, Duddalwar VA, Matcuk GR, Jr, Cen SY. Spatial assessments in texture analysis: what the radiologist needs to know. Front Radiol. (2023) 3:3. doi: 10.3389/fradi.2023.1240544, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Kaplan RC, Tirschwell DL, Longstreth WT, Jr, Manolio TA, Heckbert SR, Lefkowitz D, et al. Vascular events, mortality, and preventive therapy following ischemic stroke in the elderly. Neurology. (2005) 65:835–42. doi: 10.1212/01.wnl.0000176058.09848.bb, [DOI] [PubMed] [Google Scholar]
- 30.Homocysteine Studies Collaboration . Homocysteine and risk of ischemic heart disease and stroke. JAMA. (2002) 288:2015. doi: 10.1001/jama.288.16.2015 [DOI] [PubMed] [Google Scholar]
- 31.Cagnina A, Chabot O, Davin L, Lempereur M, Maréchal P, Oury C. Atherosclerosis, an inflammatory disease. Rev Med Liege. (2022) 77:302–9. [PubMed] [Google Scholar]
- 32.Yang Y, Wu C, Hsu P, Yang YL, Wu CH, Hsu PF, et al. Systemic immune-inflammation index (SII) predicted clinical outcome in patients with coronary artery disease. Eur J Clin Investig. (2020) 50:e13230. doi: 10.1111/eci.13230, [DOI] [PubMed] [Google Scholar]
- 33.Iadecola C, Anrather J. The immunology of stroke: from mechanisms to translation. Nat Med. (2011) 17:796–808. doi: 10.1038/nm.2399, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Anrather J, Iadecola C. Inflammation and stroke: an overview. Neurotherapeutics. (2016) 13:661–70. doi: 10.1007/s13311-016-0483-x, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Gill D, Veltkamp R. Dynamics of T cell responses after stroke. Curr Opin Pharmacol. (2016) 26:26–32. doi: 10.1016/j.coph.2015.09.009, [DOI] [PubMed] [Google Scholar]
- 36.Denorme F, Manne BK, Portier I, Eustes AS, Kosaka Y, Kile BT, et al. Platelet necrosis mediates ischemic stroke outcome in mice. Blood. (2020) 135:429–40. doi: 10.1182/blood.2019002124, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Shu JE, Ying ML, Chen XR, Hua JJ, Fu JT, Xia XM, et al. Prognostic value of high-resolution magnetic resonance imaging in evaluating carotid atherosclerotic plaque in patients with ischemic stroke. Medicine. (2017) 96:e8515. doi: 10.1097/md.0000000000008515, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. (2016) 278:563–77. doi: 10.1148/radiol.2015151169 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Rakebrandt F, Crawford DC, Havard D, Coleman D, Woodcock JP. Relationship between ultrasound texture classification images and histology of atherosclerotic plaque. Ultrasound Med Biol. (2000) 26:1393–402. doi: 10.1016/s0301-5629(00)00314-8, [DOI] [PubMed] [Google Scholar]
- 40.Baeßler B, Mannil M, Maintz D, Alkadhi H, Manka R. Texture analysis and machine learning of non-contrast T1-weighted MR images in patients with hypertrophic cardiomyopathy—preliminary results. Eur J Radiol. (2018) 102:61–7. doi: 10.1016/j.ejrad.2018.03.013, [DOI] [PubMed] [Google Scholar]
- 41.Furie KL, Kasner SE, Adams RJ, Albers GW, Bush RL, Fagan SC, et al. Guidelines for the prevention of stroke in patients with stroke or transient ischemic attack. Stroke. (2011) 42:227–76. doi: 10.1161/str.0b013e3181f7d043, [DOI] [PubMed] [Google Scholar]
- 42.International Stroke Trial Collaborative Group . The international stroke trial (IST): a randomised trial of aspirin, subcutaneous heparin, both, or neither among 19435 patients with acute ischaemic stroke. International stroke trial collaborative group. Lancet. (1997) 349:1569–81. doi: 10.1016/S0140-6736(97)04011-7, [DOI] [PubMed] [Google Scholar]
- 43.Shorr AF, Jackson WL, Moores LK. Response. Chest. (2008) 134:466–7. doi: 10.1378/chest.08-0895, [DOI] [PubMed] [Google Scholar]
- 44.Counsell C, Dennis M, McDowall M, Warlow C. Predicting outcome after acute and subacute stroke. Stroke. (2002) 33:1041–7. doi: 10.1161/hs0402.105909, [DOI] [PubMed] [Google Scholar]
- 45.Ntaios G, Faouzi M, Ferrari J, Lang W, Vemmos K, Michel P. An integer-based score to predict functional outcome in acute ischemic stroke. Neurology. (2012) 78:1916–22. doi: 10.1212/wnl.0b013e318259e221, [DOI] [PubMed] [Google Scholar]
- 46.Hall AN, Weaver B, Liotta E, Maas MB, Faigle R, Mroczek DK, et al. Identifying modifiable predictors of patient outcomes after intracerebral hemorrhage with machine learning. Neurocrit Care. (2020) 34:73–84. doi: 10.1007/s12028-020-00982-8, [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Alaka SA, Menon BK, Brobbey A, Williamson T, Goyal M, Demchuk AM, et al. Functional outcome prediction in ischemic stroke: A comparison of machine learning algorithms and regression models. Front Neurol. (2020) 11:11. doi: 10.3389/fneur.2020.00889, [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
ROC curves of the clinical model in the training and testing sets. Area under the curve (AUC).
DCA of clinical features, radiomic features, and radiomic nomogram. This figure illustrates the DCA results of clinical features, radiomic features, and radiomic nomograms, comparing the efficacy of these models in clinical decision-making. In comparison to scenarios without using any predictive models (all-treatment or no-treatment strategies), a radiomic nomogram demonstrates significant intervention benefits in patients with predictive probabilities, as opposed to clinical features and radiomic features. DCA stands for decision curve analysis.
Data Availability Statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.