Abstract
Background
Postoperative acute ischemic stroke remains a critical complication of coronary artery bypass grafting. This study aimed to develop a novel Total Cerebral Atherosclerosis Burden (TCAB) score for predicting the risk of AIS post‐coronary artery bypass grafting.
Methods
A prospective cohort of patients undergoing coronary artery bypass grafting was enrolled. The TCAB score was calculated by summing stenosis severity grades (0: <50%, 1: 50–69%, 2: 70–99%, 3: 100%) across all intracranial and extracranial artery segments. Primary outcome was in‐hospital ischemic stroke. Multivariable logistic regression models adjusted for key clinical covariates were used to evaluate the association between TCAB and clinical outcomes.
Results
Among 909 included patients, the mean TCAB score was significantly higher in patients with in‐hospital ischemic stroke compared with those without (8 versus 2, P < 0.001). A TCAB score >3 predicted in‐hospital ischemic stroke with an area under the curve of 0.756. Across all 3 multivariable models, higher TCAB scores remained independently associated with in‐hospital ischemic stroke (Model 3: odds ratio [OR]=1.089, P=0.011), 1‐year ischemic stroke (OR=1.093, P=0.011), and 1‐year major adverse cardiovascular and cerebrovascular events (OR=1.068, P=0.020). The gradient boosting machine achieved the most stable predictive ability (area under the curve=0.8736 for in‐hospital ischemic stroke; 0.8575 for 1‐year ischemic stroke; 0.7475 for 1‐year major adverse cardiovascular and cerebrovascular events).
Conclusions
The TCAB score, enhanced by machine learning, effectively predicted in‐hospital ischemic stroke, 1‐year ischemic stroke, and 1‐year major adverse cardiovascular and cerebrovascular events post‐coronary artery bypass grafting. It offers a practical tool for guiding preoperative revascularization and intraoperative embolic protection.
Keywords: acute ischemic stroke, atherosclerotic burden, coronary artery bypass grafting, machine learning, risk stratification
Subject Categories: Cerebrovascular Disease/Stroke, Ischemic Stroke
Nonstandard Abbreviations and Acronyms
- CABG
coronary artery bypass grafting
- CTA
computed tomography angiography
- GBM
gradient boosting machine
- MACCE
major adverse cardiovascular and cerebrovascular events
- TCAB
Total Cerebral Atherosclerosis Burden
- XGBoost
extreme gradient boosting
Clinical Perspective.
What Is New?
We developed a novel Total Cerebral Atherosclerosis Burden score to quantify multivessel atherosclerotic burden in patients with coronary artery bypass grafting.
What Are the Clinical Implications?
Higher Total Cerebral Atherosclerosis Burden scores were independently associated with in‐hospital ischemic stroke, 1‐year ischemic stroke, and 1‐year major adverse cardiovascular and cerebrovascular events after coronary artery bypass grafting.
The Total Cerebral Atherosclerosis Burden score, particularly when enhanced with machine learning, provides a practical and cost‐effective tool for preoperative risk stratification.
Coronary artery bypass grafting (CABG) is recommended for patients with complex coronary anatomy, such as 3‐vessel disease, particularly in those with diabetes or severe left main coronary artery stenosis. 1 However, postoperative acute ischemic stroke remains a significant complication that threatens patient outcomes. 2 , 3 , 4 Carotid artery stenosis is a well‐established risk factor for ischemic stroke following CABG. 2 , 5 , 6 Studies have shown that the incidence of perioperative ischemic stroke in patients with CABG and moderate‐to‐severe carotid artery stenosis ranges from 3% to 11%, significantly increasing the risk of mortality and long‐term disability. 7 Traditional risk assessments focus primarily on clinical factors, such as hypertension and diabetes, failing to integrate multidimensional information about atherosclerotic morphology and distribution. 8 This omission complicates accurate identification of high‐risk individuals.
Recent advancements in imaging technologies have led to the development of novel methods for quantifying the atherosclerotic burden. Lee et al. 9 demonstrated that intracranial and extracranial atherosclerosis scores measured using magnetic resonance angiography could predict ischemic stroke risk after CABG (OR=1.35). However, magnetic resonance angiography is costly, time‐consuming, and prone to overestimating the stenosis severity, which limits its clinical applicability. Kong et al. 10 further developed the Total Atherosclerotic Burden score, which combines computed tomography angiography (CTA) to assess atherosclerosis in intracranial, carotid, aortic, and coronary arteries. Their research found that this approach significantly outperformed single vascular region analysis in predicting vascular events within 1 year. However, the study had a small sample size (n=395) and did not include long‐term follow‐up data for patients undergoing CABG.
Despite these advances, most existing validated prediction models either focus on clinical variables alone or rely on advanced imaging (magnetic resonance angiography/CTA) that can be costly, time‐consuming, and not always feasible for critically ill patients or for those with claustrophobia, or metallic implants. These limitations restrict the applicability of existing models in clinical practice, underscoring the need for a practical and widely accessible risk stratification tool. Thus, this study aimed to develop an atherosclerosis burden score. Ultrasound is a quick, easy, affordable, safe, and noninvasive method, making it the preferred choice for screening vascular diseases. Studies suggest that Ultrasound screening may help detect intracranial atherosclerotic disease, which could be an independent risk factor for central nervous system complications after CABG surgery. 11 Therefore, this score will use an integrated head and neck ultrasound system to quantify atherosclerotic burden in both intracranial and extracranial arteries, combining these data with machine learning algorithms to improve prediction accuracy. The study had two primary objectives: (1) to evaluate whether a multi‐vascular region approach can predict both short‐term and long‐term ischemic stroke risk after CABG, and (2) to assess whether machine learning techniques can enhance the precision and reliability of risk assessments.
Method
The study protocol was approved by the ethics committee of Fuwai Hospital (2023–1952) of the Chinese Academy of Medical Sciences and was conducted according to the principles outlined in the Declaration of Helsinki. Written informed consent for CABG was obtained from all patients or their legally authorized representatives. As all patient information was de‐identified and anonymized before being released to the researchers for analysis, our institutional review board waived the requirement for informed consent. To obtain the data sets analyzed in this study, interested parties must submit a formal request. This request should be addressed by the corresponding author of this study.
Patients and Collection of Data
The study was performed at a single center and prospectively enrolled consecutive patients who underwent CABG and cerebral artery ultrasound and coronary CT angiography between January 2023 and January 2024 at Fuwai Hospital.
We retrieved demographic, clinical, and imaging data, including age and sex, body mass index, comorbid conditions such as previous ischemic stroke, history of hypertension, diabetes, chronic kidney disease, hyperlipidemia, and peripheral vascular disease; preoperative laboratory test results: total cholesterol (TC), triglycerides, low‐density lipoprotein (LDL) cholesterol, fasting blood glucose, creatinine, and Troponin I; surgical approach: cardiopulmonary bypass, operation time; left ventricular ejection fraction; degree of cerebral vascular stenosis and atherosclerosis burden score: intracranial atherosclerosis burden, extracranial atherosclerosis burden, Total Cerebral Atherosclerosis Burden (TCAB) based on ultrasound, and ascending aorta atherosclerosis severity based on coronary CT angiography.
Anesthesia and Surgical Procedures
Midazolam, remifentanil, and rocuronium were administered intravenously and isoflurane was used to maintain sedation levels. The patients were managed according to established practice guidelines, and all procedures were conducted using standard CABG techniques. 12 Whenever feasible, the internal thoracic artery was preferred for the revascularization of the left anterior descending artery. The decision to use cardiopulmonary bypass or specific graft materials, including arterial or saphenous vein grafts, was made at the discretion of the lead surgeon. Secondary prevention medications were recommended for all patients meeting the criteria outlined in the latest guidelines at the time of recruitment (2013 European Society of Cardiology guidelines or 2015 American Heart Association scientific statement). 13 , 14
Protocol for Cerebral Artery Ultrasound and Coronary CT Angiography
We utilized Philips EPIQ5 (Philips, Bothell, Washington) ultrasound systems equipped with 3 to 9 MHz linear array probes and 1 to 5 MHz curvilinear probes to examine the extracranial arteries, including the common carotid artery, internal carotid artery, vertebral artery, and subclavian arteries. For intracranial arteries, such as the middle cerebral arteries, terminal internal carotid arteries, anterior cerebral arteries, posterior cerebral arteries, intracranial vertebral arteries, and basilar artery, we conducted transcranial color‐coded sonography with a 1 to 5 MHz phased array probe. Peak systolic velocity and end‐diastolic velocity of arteries were recorded. According to established ultrasound criteria, the intracranial arteries include 11 segments (bilateral intracranial carotid, intracranial vertebral, anterior cerebral, middle cerebral, posterior cerebral arteries, and basilar artery). 9 , 10 The extracranial arteries included 12 segments (bilateral common carotid, carotid bulb, and internal carotid artery proximal, internal carotid artery distal, extracranial vertebral arteries v1 segments, and extracranial vertebral arteries v2‐v3 segments). 10 , 12 We defined the degree of stenosis as <50% as a score of 0, 50%–69% as 1, 70%–99% as 2, and 100% as 3 (File S1). Vertebral artery hypoplasia was not considered to reflect the stenosis. Atherosclerosis score was defined as the sum of the scores for each artery. The total score for all artery segments was 69 points. All ultrasound examinations were performed by a senior vascular ultrasonographer with more than 20 years of clinical experience, who performs approximately 3000 vascular ultrasound examinations annually and holds certification from the American Registry for Diagnostic Medical Sonography. 15 , 16 Standardized imaging protocols and diagnostic criteria were strictly followed in accordance with international guidelines to ensure consistency and minimize inter‐operator variability.
Details of the coronary CTA procedure are outlined in the supplemental material (File S2). Coronary CTA acquisition adhered to the guidelines established by the Society of Cardiovascular Computed Tomography. All coronary CTA images were anonymously transmitted to a dedicated workstation (Coronary Plaque Analysis 5.0.3, Frontier, Siemens Healthcare, Germany) 17 for analysis by 1 of the 2 trained radiologists (W.L. and Y.Z.). We divided the severity of atherosclerotic lesions in the ascending aorta into 5 grades. We defined the degree of ascending aortic atherosclerosis severity as intimal thickness <2 mm as a score of 0, focal or diffuse intimal thickness 2 to 3 mm as a score of 1 scores, 3 to 5 mm (no mobile/ulcerated components) as a score of 2, >5 mm (no mobile/ulcerated components) as a score of 3, complex grade 2 to 4 atheroma plus mobile or ulcer as a score of 4 scores. 18
Outcomes
The primary outcomes were in‐hospital ischemic stroke after CABG, which was diagnosed as (1) newly developed neurological deficits within 14 days of CABG and (2) high‐signal lesions on postoperative diffusion‐weighted imaging or low‐density lesions on postoperative CT that were not observed pre‐operatively.
Secondary outcomes were 1‐year ischemic stroke and 1‐year major adverse cardiovascular and cerebrovascular events (MACCE) occurring from hospital discharge up to 1 year. Death was defined as all‐cause mortality; ischemic stroke was defined as a cerebrovascular event characterized by focal neurological deficits, supported by either imaging evidence or a hospital diagnosis. The diagnosis of MI was based on at least 2 of the following criteria: serum creatine kinase‐myocardial band or troponin T level more than twice the upper limit of normal, new pathologic Q‐wave, history of chest discomfort for at least 30 minutes, or new left bundle branch block. Composite MACCE was defined as the occurrence of at least one of the following: myocardial infarction, cerebral infarction, or death. Y.Z., G.Z., and Y.F. completed telephone follow‐up.
Statistical Analysis Model Development
A normality test was conducted on the continuous variables and expressed as mean±SD. For non‐normally distributed data, the Wilcoxon rank‐sum test was applied, and the results were expressed as medians with interquartile ranges. Categorical variables were analyzed using Chi‐square or Fisher exact tests and presented as absolute numbers with percentages. Multiple imputation using chained equations was performed for variables with missing data constituting less than 30% of the total observations, with Troponin I being the most affected. As detailed in the Table S1 and Figure S1. To avoid overfitting while ensuring adequate adjustment for clinically relevant confounders, we constructed 3 hierarchical multivariable models: Model 1: adjusted for sex, age, and the TCAB score. Model 2: adjusted for sex, age, previous ischemic stroke, hypertension, diabetes, and the TCAB score. Model 3: adjusted for sex, age, previous ischemic stroke, hypertension, diabetes, off‐pump surgery, ascending aorta atherosclerosis severity, aspirin use, statin use, and the TCAB score. The TCAB score was modeled primarily as a continuous variable, and we additionally examined predefined TCAB categories (0, 1–3, 4–7, ≥8) to check for non‐linearity; the findings were consistent. Intracranial atherosclerosis burden, extracranial atherosclerosis burden, and TCAB were also evaluated using receiver operating characteristic curves to predict in‐hospital stroke, 1‐year stroke, and 1‐year MACCE. Kaplan–Meier curves were constructed using the categorical TCAB groups. Stratified analyses were performed based on sex (men, women), age (<70 years or ≥70 years), previous ischemic stroke, hypertension, diabetes, and LDL‐C (<1.8 mmol/L or ≥1.8 mmol/L).
Owing to the class imbalance in the dependent variables, undersampling and the synthetic minority oversampling technique (SMOTE) were applied only within the training folds during 10‐fold cross‐validation to balance the data set, while the validation folds were kept completely independent to avoid data leakage and inflated performance. Variables with a P ≤ 0.05 in the univariate analysis were incorporated into 6 machine learning models for further evaluation to predict 1‐year ischemic stroke and 1‐year MACCE, including extreme gradient boosting (XGBoost), support vector machine, adaptive boosting (AdaBoost), Naive Bayes, logistic regression, and gradient boosting machine (GBM). Gradient boosting methods (GBM, XGBoost, AdaBoost) have demonstrated high predictive performance in outcome modeling, particularly when handling complex nonlinear relationships and interactions. Support vector machine offers robust classification in high‐dimensional data spaces. Naive Bayes provides a simple probabilistic baseline that is computationally efficient and interpretable. Logistic regression remains the most widely used clinical prediction model, serving as a benchmark for evaluating potential incremental value from more complex algorithms. To ensure model robustness, 10‐fold cross‐validation combined with grid‐search optimization was used to identify the optimal hyperparameters for each algorithm. During the parameter‐tuning process, the model with the highest area under the receiver operating characteristic curve was selected as the optimal model. The final models were trained using the training set. All statistical analyses were conducted using SPSS software (version 22.0, IBM Corporation, United States) and R software (version 4.3.1, R Foundation for Statistical Computing, Austria), with a significance level set at P<0.05. YP completed all statistical analyses.
Results
This study enrolled 909 patients who underwent CABG, cerebral artery ultrasound, and coronary CT angiography between January 2023 and January 2024. Of these, 782 patients (86.1%) successfully completed the 1‐year follow‐up (Figure 1, Tables S1 and S2). A total of 21 patients (2.3%) experienced in‐hospital ischemic stroke, which was the primary outcome of the study. The baseline clinical characteristics and atherosclerosis burden scores of patients with and without in‐hospital ischemic stroke are summarized in Table 1. The mean age of the cohort was 62±9 years, with 19.5% (177/909) of the patients being women. Hypertension, diabetes, and hyperlipidemia were the most common comorbidities. Furthermore, most patients received preoperative antiplatelet and statin therapies. Patients who experienced in‐hospital ischemic stroke had a significantly higher prevalence of prior ischemic stroke (28.6% versus 10.0%, P=0.006) and anticoagulant use (28.6% versus 8.2%, P=0.001) than those who did not. Additionally, in‐hospital ischemic stroke group exhibited significantly elevated preoperative levels of total cholesterol (4 versus 3.5 mmol/L, P=0.006), LDL‐C (2.3 versus 1.9 mmol/L, P=0.011), and Troponin I (0.067 versus 0.061 ng/mL, P=0.039). The atherosclerosis burden scores, including extracranial atherosclerosis burden (6 versus 1, P<0.001), intracranial atherosclerosis burden (4 versus 1, P=0.003), and TCAB (8 versus 2, P<0.001), were substantially higher in patients the in‐hospital ischemic stroke group (Figure 2).
Figure 1. Flowchart of patient selection.
This diagram illustrates the inclusion and exclusion process for the study cohort. A total of 909 patients undergoing CABG between January 2022 and January 2023 were enrolled. Among them, 782 patients (86.1%) completed 1‐year follow‐up. Primary outcome was in‐hospital ischemic stroke, and secondary outcomes were 1‐year ischemic stroke and 1‐year MACCE. CABG indicates coronary artery bypass grafting; and MACCE, major adverse cardiovascular and cerebrovascular events.
Table 1.
Baseline Characteristics and Atherosclerosis Burden Scores in Patients With or Without Postoperative Hospital Ischemic Stroke
| Characteristics | Total (n=909) | Hospital ischemic stroke (+) (n=21) | Hospital ischemic stroke (−) (n=888) | P value |
|---|---|---|---|---|
| Age, y | 62±9 | 65±6 | 62±9 | 0.147 |
| Women, (%) | 177 (19.5) | 7 (33.3) | 170 (19.1) | 0.105 |
| BMI | 26.1±3.3 | 24.9±3.2 | 26.1±3.3 | 0.124 |
| Medical history | ||||
| Previous ischemic stroke, (%) | 95 (10.5) | 6 (28.6) | 89 (10.0) | 0.006* |
| Hypertension, (%) | 585 (64.4) | 8 (38.1) | 398 (44.9) | 0.165 |
| Diabetes, (%) | 406 (44.8) | 13 (54.2) | 45 (52.3) | 0.534 |
| Chronic kidney disease, (%) | 11 (1.2) | 1 (4.8) | 10 (1.1) | 0.228 |
| Hyperlipidemia, (%) | 614 (67.5) | 13 (61.9) | 601 (67.7) | 0.576 |
| Peripheral vascular disease, (%) | 45 (5.0) | 1 (4.8) | 44 (5.0) | 1.000 |
| Previous medication history | ||||
| Aspirin, (%) | 909 (100.0) | 21 (100.0) | 888 (100.0) | — |
| Clopidogrel, (%) | 17 (1.9) | 1 (4.8) | 16 (1.8) | 0.330 |
| Ticagrelor, (%) | 208 (22.9) | 5 (23.8) | 203 (22.9) | 1.000 |
| Anticoagulants, (%) | 79 (8.7) | 6 (28.6) | 73 (8.2) | 0.001* |
| Statins, (%) | 905 (99.6) | 21 (100.0) | 884 (99.5) | 0.758 |
| Preoperative laboratory tests | ||||
| TC, mmol/L | 3.5±1.1 | 4.0±0.9 | 3.5±1.1 | 0.006* |
| Triglycerides, mmol/L | 1.4±0.9 | 1.6±0.8 | 1.4±0.9 | 0.086 |
| LDL, mmol/L | 1.9±0.9 | 2.3±0.8 | 1.9±0.9 | 0.011* |
| FBG, mmol/L | 6.1±1.8 | 5.7±1.6 | 6.1±1.8 | 0.118 |
| Creatinine, umol/L | 73.8±22.2 | 75.7±30.1 | 73.6±21.8 | 0.740 |
| Tnl, ug/L | 0.068±0.588 | 0.061±0.137 | 0.067±0.589 | 0.039* |
| Operative data | ||||
| Operative time, (h) | 3.8 (3.2, 4.3) | 4.1 (3.2, 4.7) | 3.8 (3.2, 4.3) | 0.136 |
| Off‐pump, (%) | 399 (44.6) | 12 (57.1) | 387 (44.3) | 0.241 |
| EF, (%) | 60 (55, 65) | 61 (49, 68) | 60 (55, 65) | 0.951 |
| AAAS | 0 (0, 0) | 0 (0, 0) | 0 (0, 0) | 0.500 |
| Degree of vascular stenosis | 0.001* | |||
| <50%, (%) | 487 (53.6) | 3 (14.3) | 484 (54.5) | |
| 50%–70%, (%) | 161 (17.7) | 4 (19.0) | 157 (17.7) | |
| 70%–99%, (%) | 128 (14.1) | 8 (38.1) | 120 (13.5) | |
| 100%, (%) | 133 (14.6) | 6 (28.6) | 127 (14.3) | |
| Atherosclerosis burden score | ||||
| IAB score | 1 (0, 4) | 4 (1, 8) | 1 (0, 3) | 0.000* |
| EAB score | 1 (0, 4) | 4 (1, 8) | 1 (0, 4) | 0.003* |
| TCAB score | 3 (0, 7) | 8 (5, 12) | 2 (0, 7) | 0.000* |
Data are presented as n (%) for categorical variables, mean±SD for normally distributed continuous variables (age, BMI, and preoperative laboratory tests), and median (interquartile range) for non‐normally distributed continuous variables (operative time, EF, AAAS, and atherosclerosis burden scores).
AAAS indicates ascending aortic atherosclerosis severity; BMI, body mass index; EAB, extracranial atherosclerosis burden; EF, ejection fraction; FBG, fasting blood glucose; IAB, intracranial atherosclerosis burden; LDL, low‐density lipoprotein; TAB, total atherosclerosis burden; TC, total cholesterol; TCAB, total cerebral atherosclerosis burden; and Tnl, troponin I.
Statistically significant difference at P<0.05.
Figure 2. Distribution of TCAB score in patients with or without postoperative in‐hospital ischemic stroke.
Boxplots show the distribution of TCAB scores in patients who developed in‐hospital ischemic stroke (n=21) compared with those without stroke (n=888). Patients with in‐hospital ischemic stroke had significantly higher TCAB scores (median 8 vs 2, P<0.001). IAB indicates intracranial atherosclerosis burden; EAB, extracranial atherosclerosis burden; and TCAB, total cerebral atherosclerosis burden.
In the adjusted model (Model 3), higher TCAB scores were independently associated with in‐hospital ischemic stroke (OR=1.089, 95% CI 1.019–1.164, P=0.011) (Table 2). In contrast, ascending aorta atherosclerosis severity was not significantly associated with in‐hospital stroke after adjustment. (Table 2).
Table 2.
Multivariate Analysis of TCAB Score and Clinical Outcomes
| Model 1 | Model 2 | Model 3 | ||||
|---|---|---|---|---|---|---|
| OR (95% CI) | P value | OR (95% CI) | P value | OR (95% CI) | P value | |
| Hospital ischemic stroke | 1.099 (1.031–1.172) | 0.004* | 1.089 (1.020–1.163) | 0.011* | 1.089 (1.019–1.164) | 0.011* |
| 1‐y ischemic stroke | 1.100 (1.030–1.174) | 0.004* | 1.092 (1.021–1.168) | 0.010* | 1.093 (1.021–1.170) | 0.011* |
| 1‐y MACCE | 1.076 (1.020–1.134) | 0.007* | 1.068 (1.012–1.127) | 0.016* | 1.068 (1.010–1.129) | 0.020* |
Model 1: adjusted to sex, age, and TCAB. Model 2: adjusted to sex, age, previous ischemic stroke, hypertension, diabetes, and TCAB. Model 3: adjusted for sex, age, previous ischemic stroke, hypertension, diabetes, off‐pump, AAAS, use of aspirin and statins and TCAB.
AAAS indicates ascending aorta atherosclerosis severity; MACCE, major adverse cardiovascular and cerebrovascular events; and TCAB, total cerebral atherosclerosis burden.
Statistically significant difference at P<0.05.
Exploratory subgroup analyses for in‐hospital ischemic stroke showed no statistically significant interactions between TCAB score and sex, age, previous ischemic stroke, hypertension, diabetes, off‐pump surgery, or LDL‐C level (all P for interaction >0.05; Figures S1 and S2). Among patients without a prior history of ischemic stroke, TCAB scores were independently associated with hospital ischemic stroke (OR=1.091, 95% CI [1.011–1.176]; P=0.025). Similarly, among patients with LDL‐C levels <1.8 mmol/L, TCAB score remained a significant predictor of in‐hospital ischemic stroke (OR=1.191, 95% CI [1.045–1.358], P=0.009). There were no statistically significant differences in prior history of ischemic stroke and LDL‐C levels ≥1.8 mmol/L group.
Among the 782 patients who completed the 1‐year follow‐up, new‐onset ischemic stroke after discharge was reported in 2% (19/782), MI occurred in 1.2% (11/782), and mortality was 0.5% (5/782). In Model 3, higher TCAB scores remained independently associated with both 1‐year ischemic stroke (OR=1.093, 95% CI 1.021–1.170, P=0.011) and 1‐year MACCE (OR=1.068, 95% CI 1.010–1.129, P=0.020). (Table 2).
The predictive values of the intracranial atherosclerosis burden, extracranial atherosclerosis burden, and TCAB scores for in‐hospital ischemic stroke, 1‐year ischemic stroke, and 1‐year MACCE were evaluated (Figure 3). For in‐hospital ischemic stroke, TCAB score >3 achieved an area under the curve (AUC) of 0.756, with a sensitivity of 85.7% and specificity of 56.6%. For 1‐year ischemic stroke, TCAB score >3 yielded an AUC of 0.756, with a sensitivity of 89.5% and specificity of 56.7%. For 1‐year MACCE with a TCAB score >4, the AUC was 0.667, with a sensitivity of 62.9% and specificity of 65.6%.
Figure 3. Prediction of EAB, IAB, and TCAB score on in‐hospital ischemic stroke (A), 1‐year ischemic stroke, (B) and 1‐year MACCE (C).
ROC curves show the ability of IAB, EAB, and TCAB to predict: (A) in‐hospital ischemic stroke, (B) 1‐year ischemic stroke, and (C) 1‐year MACCE. For in‐hospital ischemic stroke, TCAB >3 achieved AUC=0.756, sensitivity 85.7%, and specificity 56.6%. For 1‐year ischemic stroke, TCAB >3 achieved AUC=0.756, sensitivity 89.5%, specificity 56.7%. For 1‐year MACCE, TCAB >4 achieved AUC=0.667. IAB indicates intracranial atherosclerosis burden; EAB, extracranial atherosclerosis burden; TCAB, total cerebral atherosclerosis burden; MACCE, major adverse cardiovascular and cerebrovascular events; ROC, receiver operating characteristic; and AUC, area under the curve.
Figure S3 illustrates the survival analysis based on TCAB stratification into very low‐risk (0 score), low‐risk (1–3 score), medium‐risk (4–7 score), and high‐risk groups (≥8 score). Compared with the very low‐risk group, patients in the high‐risk group demonstrated significantly higher rates of 1‐year ischemic stroke and 1‐year MACCE (P<0.05). However, there were no significant statistical differences between the very low‐ and low‐risk groups and the very low‐ and medium‐risk groups.
The use of 6 machine learning models to predict in‐hospital ischemic stroke was explored: AdaBoost, GBM, logistic regression, Naive Bayes, SVM, and XGBoost. In predictive performance analysis, the GBM model demonstrated the most stable predictive ability. For in‐hospital ischemic stroke and 1‐year ischemic stroke, GBM achieved an AUC of 0.8736 and 0.8575 significantly higher than TCAB alone (P<0.05). For 1‐year MACCE, GBM achieved an AUC of 0.7475 compared with 0.667 for TCAB, showing a non‐significant trend toward improvement (P>0.05) (Figure 4, Figure S4).
Figure 4. Prediction of in‐hospital ischemic stroke (A), 1‐year ischemic stroke (B), and 1‐year MACCE (C) using 6 machine learning models.
ROC curves compare 6 machine learning models—AdaBoost, GBM, logistic regression, Naive Bayes, SVM, and XGBoost—in predicting (A) in‐hospital ischemic stroke, (B) 1‐year ischemic stroke, and (C) 1‐year MACCE. Among these, the GBM model demonstrated the most stable predictive ability, achieving AUC=0.8736 for in‐hospital ischemic stroke, 0.8575 for 1‐year ischemic stroke, and 0.7475 for 1‐year MACCE. AdaBoost indicates adaptive boosting; AUC, area under the curve; GBM, gradient boosting machine; ROC, receiver operating characteristic; SVM, support vector machine; and XGBoost, extreme gradient boosting.
DISCUSSION
Our study demonstrated that the TCAB score integrating the assessment of atherosclerotic burden across multiple vascular regions is a significant independent predictor of clinical outcomes. In the adjusted model 3, higher TCAB scores were significantly associated with in‐hospital ischemic stroke (OR=1.089), 1‐year ischemic stroke (OR=1.093), and 1‐year MACCE (OR=1.068). Furthermore, the incorporation of machine learning algorithms, such as GBM, enhances the accuracy of risk stratification, providing clinicians with a more effective tool for assessing patient risk.
This study introduces a notable advancement in the assessment of multivascular regions. Consistent with prior research by Lee, our study identified a relationship between atherosclerosis in both the intracranial and extracranial arteries and post‐CABG ischemic stroke. 9 However, extracranial atherosclerotic burden is a stronger predictor of in‐hospital stroke. Extracranial plaques with large lipid cores and thin fibrous caps are more susceptible to embolism during surgery, whereas intracranial arteries are protected by collateral circulation. 19 , 20 , 21 In addition, our study evaluated the atherosclerotic burden of the ascending aorta, which is critically important because it is strongly associated with postoperative ischemic stroke. 22 , 23 , 24 After adjusting for multiple variables, TCAB remained significantly correlated with postoperative ischemic stroke, 1‐year ischemic stroke, and 1‐year MACCE, corroborating the findings of Lainelehto. 10 However, our study did not include data on postoperative medication adherence, which may have limited the broader applicability of these results. In addition, the significantly higher preoperative anticoagulant use in the stroke group (28.6% versus 8.2%) likely reflects confounding by indication. Patients on anticoagulants typically possess high‐risk cardioembolic substrates (e.g., atrial fibrillation or valvular disease), representing a heavier baseline risk that contributes to the excess perioperative stroke incidence.
Notably, current clinical practices vary globally: In the United States, carotid duplex ultrasound is a common pre‐CABG assessment, mainly for patients with previous stroke or high neurological risk. In contrast, this practice is less common in other regions, particularly in Asian populations where intracranial atherosclerosis is more prevalent. Epidemiological studies confirm higher rates of intracranial atherosclerotic stenosis in these groups, 25 underscoring the need for a comprehensive evaluation of both intracranial and extracranial arteries. Thus, the TCAB score—based on a unified ultrasound evaluation of both intracranial and extracranial atherosclerosis—addresses this gap and enhances predictive accuracy across diverse populations.
Regarding perioperative management, the study found that higher TCAB scores were associated with an increased risk of ischemic stroke. Patients with TCAB scores >3 experienced an ischemic stroke rate of 4.4% during surgery and 4.8% at 1 year, which was significantly higher than the 0.8%–2.0% ischemic stroke rate typically observed in CABG. 3 The TCAB score can help guide surgical decisions such as prioritizing off‐pump techniques 8 and using embolism protection strategies. 26 Furthermore, exploratory subgroup showed that even low‐risk populations with a high atherosclerotic burden may face residual risk. However, results from exploratory subgroup analyses must be interpreted with caution given the low number of events.
The integration of machine‐learning models further refined the predictive performance of TCAB for ischemic stroke and MACCE. Our study observed that the predictive power of TCAB decreased over time, suggesting that the atherosclerotic burden of a single vascular region may not fully explain the clinical outcomes. 27 By incorporating statistically significant variables from the univariate analysis into 6 different machine learning models, it was found that the GBM model provided the most stable performance. The improvement was statistically significant for 1‐year ischemic stroke, with AUCs of 0.8575 versus 0.756 for TCAB. For 1‐year MACCE, GBM yielded an AUC of 0.7475 compared with 0.667 for TCAB, suggesting a favorable trend though the difference did not reach statistical significance. In addition, the study did not assess plaque characteristics, such as density or size, which could offer additional predictive value. To further enhance predictive accuracy, future research should explore the integration of machine learning with deep learning techniques to incorporate a wider atherosclerotic parameter. 28 , 29 Given the dynamic and time‐dependent nature of the surgical process, the application of time‐series modeling that incorporates preoperative, intraoperative, and postoperative data could significantly enhance perioperative risk stratification. In our further work, we aim to build a user‐friendly, web‐based clinical decision‐support tool that embeds the ML‐enhanced TCAB score. This system will provide automated, individualized risk predictions, visual analytics, and scenario‐based recommendations, with potential for seamless integration into hospital information systems to support perioperative decision‐making.
The ultrasound‐based TCAB scoring system is a simple, inexpensive, and practical approach that can be used to predict perioperative stroke risk in patients undergoing CABG. Its integration into routine preoperative evaluation could serve as a valuable clinical decision‐support tool. However, broader implementation will require multicenter validation, simplification of the model for settings lacking advanced imaging, and strategies to overcome technical limitations such as poor temporal bone windows. This study has several limitations. Firstly, as a single‐center study with a modest sample size (n=909) and a low event rate (only 21 in‐hospital ischemic strokes and 19 1‐year strokes), the statistical power was limited, particularly for detecting interactions in exploratory subgroup analyses. These analyses require validation in larger cohorts. Secondly, the predominance of male patients in our cohort (approximately 80%) reflects the higher prevalence of coronary and carotid atherosclerosis in men but may further limit generalizability to a more sex‐balanced population. In the future, we will conduct a larger prospective study to validate our conclusions. Thirdly, the TCAB score did not incorporate data on coronary or peripheral arterial disease, and future research could expand this to include a comprehensive systemic atherosclerosis score. Fourthly, the lack of external validation for machine learning models and the absence of direct comparison with traditional clinical scores (such as EuroSCORE II) calls for further investigation into their added value. In addition, we did not perform a formal reproducibility assessment between 2 independent ultrasonographers to quantify inter‐observer variability. However, all examinations were conducted by an experienced, certified ultrasonographer using standardized protocols, which likely reduced variability and improved the reliability of the measurements. Moreover, The cutoff values were optimized using the Youden index from this data set, which may introduce optimism bias. Validation in external multicenter cohorts is needed to verify their generalizability. Fifthly, specific diagnoses of atrial fibrillation were not recorded in the anonymized data set. We therefore relied on anticoagulant use as a surrogate for cardioembolic risk, as retrospective retrieval was precluded by de‐identification protocols. Finally, while we used an ultrasound‐based scoring system, its application requires specialized training and significant clinical experience from ultrasound specialists. It limits the implementation of model. Additionally, our study excluded certain surgical patient populations with poor temporal bone windows, which limit visibility and prevent effective transcranial doppler assessments.
Conclusions
The TCAB score effectively predicted in‐hospital ischemic stroke, 1‐year ischemic stroke, and 1‐year MACCE risk following CABG. The integration of machine learning models further refines risk stratification, aiding clinical decision making for high‐risk patients. Future studies should extend the TCAB score to include systemic atherosclerosis assessment and validate it through multi‐center cohorts to optimize perioperative management.
Sources of Funding
This study was supported by the Noncommunicable Chronic Diseases‐National Science and Technology Major Project (2023ZD0504703), National High‐Level Hospital Clinical Research Funding (2022‐GSP‐TS‐4, 2023‐GSP‐GG‐37, 2023‐GSPQN‐12), Fuwai Hospital Chinese Academy of Medical Sciences Specialty Discipline Development Project (2022‐FWTS05), CAMS Innovation Fund for Medical Sciences (CIFMS) (2022‐I2M‐C&T‐B‐046), and Fundamental Research Funds for the Central Universities (3332024041).
Disclosures
All authors declare no competing interests.
Supporting information
Files S1–S2
Tables S1–S2
Figures S1–S4
STROBE_Checklist
Acknowledgments
Y.P. performed analyses and drafted the manuscript. J.L., Y.Z., and F.X. provided clinical advice and managed the significance of clinical metrics. Y.P. and G.Z. assisted in creating and revising tables and figures. WL and NL performed the imaging assessments. Y.F., N.Z., Y.M., X.L., and Q.Z. performed secondary checking of data and tables and graphs. J.L. and Y.Z. planned and supervised the study and revised the manuscript. All authors provided final approval for the version to be published.
This manuscript was sent to Fadar O. Otite, MD, SM, Associate Editor, for review by expert referees, editorial decision, and final disposition.
Supplemental Material is available at https://www.ahajournals.org/doi/suppl/10.1161/JAHA.125.043567
For Sources of Funding and Disclosures, see page 11.
Contributor Information
Fei Xu, Email: felix_xufei@126.com.
Yinghua Zhou, Email: yhzhou2008@sina.com.
Shujuan Li, Email: lishujuan@fuwai.com.
References
- 1.Writing Committee M, Lawton, J. S. , Tamis‐Holland, J. E. et al 2022. 2021 ACC/AHA/SCAI guideline for coronary artery revascularization: a report of the American College of Cardiology/American Heart Association joint Committee on clinical practice guidelines. J Am Coll Cardiol;79:e21‐e129. doi:DOI: 10.1016/j.jacc.2021.09.006 [DOI] [PubMed] [Google Scholar]
- 2. Fanning JP, Campbell BCV, Bulbulia R, Gottesman RF, Ko S‐B, Floyd TF, Messé SR. Perioperative stroke. Nat Rev Dis Primers. 2024;10(1):3. doi: 10.1038/s41572-023-00487-6 [DOI] [PubMed] [Google Scholar]
- 3. Gaudino M, Rahouma M, Di Mauro M, Yanagawa B, Abouarab A, Demetres M, Di Franco A, Arisha MJ, Ibrahim DA, Baudo M, et al. Early versus delayed stroke after cardiac surgery: a systematic review and meta‐analysis. J Am Heart Assoc. 2019;8:e012447. doi: 10.1161/JAHA.119.012447 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Vermeer SE, Prins ND, den Heijer T, Hofman A, Koudstaal PJ, Breteler MM. Silent brain infarcts and the risk of dementia and cognitive decline. N Engl J Med. 2003;348:1215–1222. doi: 10.1056/NEJMoa022066 [DOI] [PubMed] [Google Scholar]
- 5. Masabni K, Sabik JF 3rd, Raza S, et al. Nonselective carotid artery ultrasound screening in patients undergoing coronary artery bypass grafting: is it necessary? J Thorac Cardiovasc Surg. 2016;151:402–408. doi: 10.1016/j.jtcvs.2015.09.108 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Zhang H, Zhang D, Qu J, Wang J, Rao C, Chen S, Zhao Y, Li H, Gao G, Liu S, et al. Association of carotid duplex ultrasonography screening with stroke and mortality among patients undergoing coronary artery bypass grafting. J Vasc Surg. 2024;80:153–162 e4. doi: 10.1016/j.jvs.2024.02.039 [DOI] [PubMed] [Google Scholar]
- 7. Sultan I, Bianco V, Kilic A, Jovin T, Jadhav A, Jankowitz B, Aranda‐Michel E, D’angelo MP, Navid F, Wang Y, et al. Predictors and outcomes of ischemic stroke after cardiac surgery. Ann Thorac Surg. 2020;110:448–456. doi: 10.1016/j.athoracsur.2020.02.025 [DOI] [PubMed] [Google Scholar]
- 8. Tachibana H, Hiraoka A, Saito K, Naito Y, Chikazawa G, Tamura K, Totsugawa T, Yoshitaka H, Sakaguchi T. Incidence and impact of silent brain lesions after coronary artery bypass grafting. J Thorac Cardiovasc Surg. 2021;161:636–644. doi: 10.1016/j.jtcvs.2019.09.162 [DOI] [PubMed] [Google Scholar]
- 9. Lee EJ, Choi KH, Ryu JS, Jeon SB, Lee SW, Park SW, Park SJ, Lee JW, Choo SJ, Chung CH, et al. Stroke risk after coronary artery bypass graft surgery and extent of cerebral artery atherosclerosis. J Am Coll Cardiol. 2011;57:1811–1818. doi: 10.1016/j.jacc.2010.12.026 [DOI] [PubMed] [Google Scholar]
- 10. Lainelehto K, Pienimaki JP, Savilahti S, Huhtala H, Numminen H, Putaala J. Cervicocerebral atherosclerosis burden increases long‐term mortality in patients with ischemic stroke or transient ischemic attack. J Am Heart Assoc. 2024;13:e032938. doi: 10.1161/JAHA.123.032938 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Yoon BW, Bae HJ, Kang DW, Lee SH, Hong KS, Kim KB, Park BJ, Roh JK. Intracranial cerebral artery disease as a risk factor for central nervous system complications of coronary artery bypass graft surgery. Stroke. 2001;32:94–99. doi: 10.1161/01.str.32.1.94 [DOI] [PubMed] [Google Scholar]
- 12. Wang C, Hu Z, Hou Z, Wang Y, Song L, Xu B, Guan C, Ning Y, Feng W, Zhang Y. Impact of preoperative quantitative flow ratio of the left anterior descending artery on internal mammary artery graft patency and midterm patient outcomes after coronary artery bypass grafting. J Am Heart Assoc. 2023;12:e029134. doi: 10.1161/JAHA.122.029134 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Task Force M , Montalescot G, Sechtem U, et al. ESC guidelines on the management of stable coronary artery disease: the Task Force on the management of stable coronary artery disease of the European Society of Cardiology. Eur Heart J. 2013;34:2949–3003. doi: 10.1093/eurheartj/eht296 [DOI] [PubMed] [Google Scholar]
- 14. Kulik A, Ruel M, Jneid H, Ferguson TB, Hiratzka LF, Ikonomidis JS, Lopez‐Jimenez F, McNallan S, Patel M, Roger VL, et al. Secondary prevention after coronary artery bypass graft surgery: a scientific statement from the American Heart Association. Circulation. 2015;131:927–964. doi: 10.1161/CIR.0000000000000182 [DOI] [PubMed] [Google Scholar]
- 15. AIUM practice parameter for the performance of transcranial Doppler ultrasound. J Ultrasound Med. 2023;42:E36–E44. doi: 10.1002/jum.16234 [DOI] [PubMed] [Google Scholar]
- 16. Gaudino M, Benesch C, Bakaeen F, DeAnda A, Fremes SE, Glance L, Messé SR, Pandey A, Rong LQ; American Heart Association Council on Cardiovascular Surgery and Anesthesia; Stroke Council; and Council on Cardiovascular and Stroke Nursing . Considerations for reduction of risk of perioperative stroke in adult patients undergoing cardiac and thoracic aortic operations: a scientific statement from the American Heart Association. Circulation. 2020;142:e193–e209. doi: 10.1161/CIR.0000000000000885 [DOI] [PubMed] [Google Scholar]
- 17. Lu ZF, Yin WH, Schoepf UJ, Abrol S, Ma JW, Zhao L, Su XM, An YQ, Xiao ZC, Lu B. Prediction value of pericoronary fat attenuation index for coronary in‐stent restenosis. Eur Radiol. 2024;34:4950–4959. doi: 10.1007/s00330-023-10527-0 [DOI] [PubMed] [Google Scholar]
- 18. Knol WG, Budde RPJ, Mahtab EAF, Bekkers JA, Bogers A. Intimal aortic atherosclerosis in cardiac surgery: surgical strategies to prevent embolic stroke. Eur J Cardiothorac Surg. 2021;60:1259–1267. doi: 10.1093/ejcts/ezab344 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Huang Z, Cheng XQ, Lu RR, Bi XJ, Liu YN, Deng YB. Incremental prognostic value of carotid plaque‐RADS over stenosis degree in relation to stroke risk. JACC Cardiovasc Imaging. 2025;18:77–89. doi: 10.1016/j.jcmg.2024.07.004 [DOI] [PubMed] [Google Scholar]
- 20. Huang K, Yao W, Song Z, Jia X, Gao J, Liu R, Han Y, Liu X, du J, Tu S, et al. Prognostic value of angiographic based quantitative flow ratio and anatomic features in intracranial atherosclerotic stenosis. J Neurointerv Surg. 2025;17:980–985. doi: 10.1136/jnis-2024-022007 [DOI] [PubMed] [Google Scholar]
- 21. Liu X, Dornbos D 3rd, Pu Y, Leng X, Song L, Jia B, Pan Y, Wang D, Miao Z, Wang Y, et al. Collateral circulation alters downstream hemodynamic stress caused by intracranial atherosclerotic stenosis. Neurol Res. 2017;39:498–503. doi: 10.1080/01616412.2017.1315483 [DOI] [PubMed] [Google Scholar]
- 22. Park KH, Lee HY, Lim C, Chung ES, Sung SW, Choi SI, Chun EJ. Clinical impact of computerised tomographic angiography performed for preoperative evaluation before coronary artery bypass grafting. Eur J Cardiothorac Surg. 2010;37:1346–1352. doi: 10.1016/j.ejcts.2009.12.040 [DOI] [PubMed] [Google Scholar]
- 23. Chlorogiannis DD, Pargaonkar S, Apostolos A, Vythoulkas‐Biotis N, Kokkinidis DG, Nagraj S. The predictive value of aortic calcification on computed tomography for major cardiovascular events. J Clin Med. 2024;13:4019. doi: 10.3390/jcm13144019 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Lantelme P, Eltchaninoff H, Rabilloud M, Souteyrand G, Dupré M, Spaziano M, Bonnet M, Becle C, Riche B, Durand E, et al. Development of a risk score based on aortic calcification to predict 1‐year mortality after Transcatheter aortic valve replacement. JACC Cardiovasc Imaging. 2019;12:123–132. doi: 10.1016/j.jcmg.2018.03.018 [DOI] [PubMed] [Google Scholar]
- 25. Gutierrez J, Turan TN, Hoh BL, Chimowitz MI. Intracranial atherosclerotic stenosis: risk factors, diagnosis, and treatment. Lancet Neurol. 2022;21:355–368. doi: 10.1016/S1474-4422(21)00376-8 [DOI] [PubMed] [Google Scholar]
- 26. Butala NM, Kapadia SR, Secemsky EA, Gallup D, Kosinski AS, Vemulapalli S, Messenger JC, Yeh RW, Cohen DJ. Impact of cerebral embolic protection devices on disabling stroke after TAVR: updated results from the STS/ACC TVT registry. Circ Cardiovasc Interv. 2024;17:e013697. doi: 10.1161/CIRCINTERVENTIONS.123.013697 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Mahmoudi M, Hill PC, Xue Z, Torguson R, Ali G, Boyce SW, Bafi AS, Corso PJ, Waksman R. Patients with severe asymptomatic carotid artery stenosis do not have a higher risk of stroke and mortality after coronary artery bypass surgery. Stroke. 2011;42:2801–2805. doi: 10.1161/STROKEAHA.111.618082 [DOI] [PubMed] [Google Scholar]
- 28. van Dam‐Nolen DHK, Truijman MTB, van der Kolk AG, Liem MI, Schreuder FHBM, Boersma E, Daemen MJAP, Mess WH, van Oostenbrugge R, van der Steen A, et al. Carotid plaque characteristics predict recurrent ischemic stroke and TIA: the PARISK (plaque At RISK) study. JACC Cardiovasc Imaging. 2022;15:1715–1726. doi: 10.1016/j.jcmg.2022.04.003 [DOI] [PubMed] [Google Scholar]
- 29. Kopczak A, Schindler A, Sepp D, Bayer‐Karpinska A, Malik R, Koch ML, Zeller J, Strecker C, Janowitz D, Wollenweber FA, et al. Complicated carotid artery plaques and risk of recurrent ischemic stroke or TIA. J Am Coll Cardiol. 2022;79:2189–2199. doi: 10.1016/j.jacc.2022.03.376 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Files S1–S2
Tables S1–S2
Figures S1–S4
STROBE_Checklist