Opportunistic Cardiovascular Risk Assessment Using Routine Head CT in the Emergency Department

Abstract

BACKGROUND

Routine noncardiac computed tomography (CT) imaging may contain information about cardiovascular risk. Head computed tomography (CTH) is among the most common imaging studies, conducted annually in millions of patients. Its utility for cardiovascular risk assessment has not been studied.

OBJECTIVES

The purpose of this study was to develop and validate deep learning models for predicting incident cardiovascular disease (CVD) and estimating coronary artery calcium (CAC) scores from CTH, and assess performance against clinical risk factors.

METHODS

This retrospective cohort study used data from the Stanford Health Care Emergency Department from August 2020 to August 2024. The CVD cohort comprised 27,990 adult patients without known CVD who underwent CTH. The CAC cohort included 2,313 patients who underwent both CTH and coronary CT angiography. Imaging features were extracted from CTH using pretrained deep learning models. Other risk factors were extracted from electronic health records. Outcomes were incident CVD complications (myocardial infarction, stroke, heart failure) and CAC scores (0, 1–10, 11–100, 101–400, >400 AU). Performance was evaluated using the concordance index (C-index) and area under the receiver operating characteristic curve, compared against the baseline model using the variables of the American Heart Association PREVENT (Predicting Risk of cardiovascular disease EVENTs) risk model.

RESULTS

Four percent (1,110 of 27,990) of patients (median age 63.0 years, 51.7% female) experienced cardiovascular events. The CTH model achieved a C-index of 0.82 (95% CI: 0.78–0.85) compared with PREVENT (0.75; 95% CI: 0.70–0.79) with difference of 0.07 (95% CI: 0.04–0.10). For CAC estimation (n = 2,313, median age 65.0 years, 53.5% female), the CTH+PREVENT model achieved a C-index of 0.76 (95% CI: 0.72–0.80) and area under the receiver operating characteristic curve of 0.80 (95% CI: 0.73–0.85) for CAC >100. Of the patients, 15.7% were reclassified; higher-risk patients were younger but with higher prevalence of vascular calcifications (30.2% vs 24.8%, P = 0.001) and brain infarcts (20.1% vs 5.8%, P < 0.001).

CONCLUSIONS

Routine CTH scans complement traditional risk factors for cardiovascular risk stratification, identifying subclinical disease in younger patients with favorable risk profiles. Clinical integration could improve CVD detection and prevention without additional costs or radiation.

CENTRAL ILLUSTRATION

Complementary Cardiovascular Risk Assessment From Routine Head Computed Tomography

Traditional guideline-based clinical risk factors (PREVENT framework) are integrated with imaging-derived features extracted from routine emergency department CTH scans. CTH provides complementary information reflecting subclinical vascular disease that is not captured by clinical risk factors alone. The integrated assessment refines cardiovascular risk stratification, enabling improved prediction of incident cardiovascular disease and enhanced identification of elevated coronary artery calcium burden, with 15.7% of patients reclassified without additional imaging, cost, or radiation exposure. AI = artificial intelligence; AUC = area under the curve; CAC = coronary artery calcium; CT = computed tomography; CTH = head computed tomography; CVD = cardiovascular disease; eGFR = estimated glomerular filtration rate; HDL = high-density lipoprotein; PREVENT = Predicting Risk of cardiovascular disease EVENTs; SBP = systolic blood pressure.

Cardiovascular disease (CVD) is the leading cause of death globally and may progress silently over years before leading to compli cations such as stroke or myocardial infarction. Early detection and treatment of high-risk individuals is the cornerstone of prevention, with traditional risk equations using demographic, clinical, and laboratory data.^1,2

Imaging can refine risk prediction. For example, coronary artery calcium (CAC) scoring from electrocardiogram-gated cardiac computed tomography (CT) scans can help identify patients who may benefit from early intervention,³ improving on clinical risk factors alone.⁴ Subclinical or “silent” infarcts on brain magnetic resonance imaging double stroke risk in older adults,⁵ suggesting that advanced imaging may contain evidence of latent cardiovascular risk.

The widespread use of advanced imaging is limited by cost and accessibility. Recent advances in artificial intelligence (AI) enable opportunistic cardiovascular risk assessment using imaging performed for unrelated clinical indications. For example, deep learning models can estimate CAC from nongated chest CT scans,^6,7 offering a scalable and cost-effective alternative to dedicated cardiac imaging. Other researchers have studied opportunistic screening from other studies, including abdominal CT and chest radiographs.^8–10

Head CT (CTH) is among the most common imaging studies, second only to chest radiographs, with an estimated 22.5 million scans conducted annually in the United States,¹¹ most in the emergency department (ED). These studies represent a potential resource for opportunistic screening,^12,13 particularly in patients who might not otherwise receive dedicated cardiac imaging. Leveraging CTH to assess cardiovascular risk could enable earlier identification of high-risk individuals without additional cost, radiation, or workflow disruption.

This study aimed to develop and validate deep learning models to predict future cardiovascular events and estimate CAC scores using features derived from routine CTH studies. The primary objectives were: 1) to develop and validate a deep learning model to predict the time to incident CVD events (myocardial infarction, stroke, or heart failure); 2) to develop and validate a deep learning model to estimate CAC scores;¹⁴ and 3) to compare the performance of CTH models, alone and in combination with established risk factors, against a clinical baseline model using the variables of the American Heart Association (AHA) PREVENT (Predicting Risk of cardiovascular disease EVENTs) risk score² (Central Illustration).

We hypothesized that deep learning models using features from CTH could predict incident CVD events and CAC scores with accuracy comparable to established clinical risk models, and provide complementary information relevant to near-term cardiovascular risk.

METHODS

STUDY DESIGN.

This study was approved by the Stanford University Institutional Review Board (IRB) (58581) and Harvard University IRB (22–0364) with waiver of consent for retrospective research on deidentified, routinely collected data.

We used a proprietary dataset consisting of 98,175 CTH studies from 35,237 patients across 3 U.S. health systems (1 urban health system in North Carolina, 1 rural health system in Missouri, and 1 multisite urban health system spanning New York and Texas) for pretraining the CTH model.¹⁵ We conducted primary analyses on data from the Stanford Health Care ED, a subset of which was published as “Multimodal Clinical Monitoring in the Emergency Department (MC-MED).”^16,17 The dataset comprises linked clinical, imaging, and longitudinal outcome data for adult patients treated at Stanford Health Care between August 1, 2020, and August 30, 2024. Data were derived from the electronic health record, radiology Picture Archiving and Communication System, and the California vital statistics database.

COHORT CONSTRUCTION AND OUTCOMES.

We constructed 2 cohorts from 35,935 adult patients with CTH performed between August 1, 2020, and August 30, 2024 (Figure 1). Baseline characteristics of the cohorts are presented in Table 1.

FIGURE 1. Flow Diagram of the Study Cohort.

The CTH pretraining dataset contains 98,175 studies from 35,237 patients. From the Stanford ED Dataset, containing 313,966 total visits from 160,016 unique patients, 35,935 patients with CTH scans were identified. This population was divided into 2 cohorts: the CVD prediction cohort (27,990 patients with no prior CVD based on past medical history or diagnoses during the index ED visit) and the CAC score prediction cohort (2,313 patients with both CTH and CCTA with available CAC scores). CAC = coronary artery calcium; CCTA = coronary computed tomography angiography; CT = computed tomography; CTH = head computed tomography; CVD = cardiovascular disease; ED = emergency department.

TABLE 1.

Cohort Demographics and Statistics for CVD Prediction and CAC Score Prediction

	CVD Prediction (n = 27,990)	CAC Score Prediction (n = 2,313)
Age, y	63.0 (44.0–77.0)	65.0 (55.0–74.0)
Female	14,481 (51.7)	1,238 (53.5)
Male	13,480 (48.2)	1,074 (46.4)
Sex: Unknown	29 (0.1)	1 (0.0)
SBP, mm Hg	132.0 (119.0–146.0)	134.0 (120.5–148.0)
Total cholesterol, mg/dL	160.0 (128.0–198.0)	162.0 (128.0–198.0)
HDL cholesterol, mg/dL	49.0 (38.0–62.0)	48.0 (38.0–63.0)
eGFR, mL/min/1.73 m2	89.0 (67.0–105.0)	85.0 (65.0–97.0)
Obesity	4,534 (16.2)	505 (21.8)
Hypertension	5,819 (20.8)	1,022 (44.2)
Diabetes	2,013 (7.2)	361 (15.6)
Antihypertensive medicine	4,928 (17.6)	610 (26.4)
Lipid-lowering medications	4,216 (15.1)	547 (23.6)
Smoking status	1,627 (8.6)	114 (5.9)

Values are median (IQR) or n (%).

CAC = coronary artery calcium; CVD = cardiovascular disease; eGFR = estimated glomerular filtration rate; HDL = high-density lipoprotein; SBP = systolic blood pressure.

CVD prediction cohort (n 27,990).

This cohort included patients without prior CVD based on past medical history or diagnoses during the index ED visit (Supplemental Methods). This exclusion ensured that the cohort consisted of patients without known CVD at the time of CTH imaging. CVD events were defined using International Classification of Diseases codes for myocardial infarction (I21, I22), stroke (I61, I62, I63), and heart failure (I50) documented during subsequent encounters after the index ED visit, consistent with AHA PREVENT definitions.² Follow-up extended from CTH until CVD event, death, or study end (August 30, 2024).

CAC score estimation cohort (n = 2,313).

This cohort included patients with both CTH and coronary computed tomography angiography (CCTA) at any time during the study period. For patients with multiple CTH studies, the CTH study closest in time to CCTA was selected. The median time difference between CTH and CCTA was − 116.4 days (IQR: − 977.3 to 33.6 days) in the training set, with 37.9% having CTH performed before CCTA. Given that CAC reflects stable atherosclerotic disease with recommended rescanning intervals of 3 to 7 years,¹⁸ the time difference of approximately 4 months in our cohort is well within the period in which CAC burden would be expected to remain stable. Radiology-reported CAC scores were extracted from CCTA reports using GPT-4.1 (2025-04-14) and categorized as 0, 1–10, 11–100, 101–400, or >400 AU based on established clinical thresholds¹⁴ (Supplemental Table 1).

THE PREVENT-CVD MODEL.

We used the AHA PREVENT-CVD model as our primary baseline model, as PREVENT represents the most current evidence-based approach to cardiovascular risk assessment.² Clinical variables were extracted from the electronic health record, using data available through the conclusion of the ED visit in which CTH was obtained. These included age, sex, high-density lipoprotein (HDL) and total cholesterol levels, estimated glomerular filtration rate, systolic blood pressure, smoking status, obesity status (substituted for body mass index), diabetes mellitus, and use of antihypertensive or lipid-lowering medications. Specific extraction methods are detailed in the Supplemental Methods.

Diagnostic codes, medication definitions, and clinical note extraction prompts are provided in Supplemental Tables 1 and 2. Missingness patterns are reported in Supplemental Table 3. Missing values were imputed using median values of the training set, with sensitivity analysis using multiple imputation (Supplemental Methods, Supplemental Tables 4–9).

MODEL DEVELOPMENT.

For both CVD event prediction and CAC score estimation, we developed 3 models: 1) PREVENT-CVD (variables from the AHA PREVENT risk equation), 2) CTH (imaging features alone), and 3) CTH+PREVENT (combined imaging and clinical features). In supplementary analyses, we evaluated the pooled cohort equations as an additional comparator (Supplemental Tables 10 and 11).

For CTH models, we pretrained a vision transformer^19,20 using a separate proprietary dataset of CTH studies. We pretrained a vision transformer using DINOv2,¹⁹ a self-supervised learning framework, on 98,175 CTH studies from 3 external health systems. This stage learned general visual representations from CTH images without requiring labels. We further trained the vision transformer using CLIP (Contrastive Language-Image Pre-training),²⁰ which learns to align CTH images with their corresponding radiology reports. Patients were 57.0% female, with a mean age of 65.7 years (IQR: 53.0–78.0) (Supplemental Figure 1). Analyses of pretraining frameworks are provided in Supplemental Tables 12 and 13. For downstream prediction tasks, we trained linear models on CTH embeddings. For CVD event prediction, we used a linear model trained with a Cox proportional-hazards loss function. For CAC estimation, we used a linear model trained with crossentropy loss as a multiclass classifier.

For clinical variables, we used a linear model with the same loss function as the corresponding CTH model. All models were trained and evaluated on 80/10/10 train/validation/test splits. Details the characteristics of each split are provided in Supplemental Tables 14 and 15.

Details on model architecture, pretraining, and hyperparameter tuning are described in the Supplemental Methods and Supplemental Figure 2.

STATISTICAL ANALYSIS.

Model performance was evaluated using the concordance index (C-index). For CVD time-to-event models, we also calculated the area under the receiver operating characteristic curve at multiple time points. For the CAC model, we calculated the area under the receiver operating characteristic curve for clinically relevant thresholds. Model calibration was assessed using the Brier score and the Integrated Calibration Index (ICI). All models were assessed for adherence to the proportional-hazards assumption (Supplemental Methods, Supplemental Figure 3). We also assessed how the addition of CTH features affected patient risk rankings (Supplemental Methods).

To assess model generalizability across time, we performed temporal validation by training models on data from August 2020 to December 2023 and testing on data from January to August 2024. This split simulates prospective deployment and evaluates whether model performance remains stable as patient populations and clinical practices evolve.

CIs were computed using bootstrap resampling with 1,000 iterations, sampling with replacement from the test set. The 95% CIs represent the 2.5 ^th and 97.5 ^th percentiles of the bootstrap distribution. Proportions were compared using chi-square tests, and medians were compared using Mann-Whitney U tests. Statistical analysis was performed using Python version 3.12.9 with scikit-learn, PyTorch,²¹ and lifelines²² packages.

RESULTS

A total of 35,935 ED patients underwent CTH during the study period. Of these, 27,990 had no history of CVD documented either in their past medical history or diagnosed during the index ED visit in which the CTH was obtained, and were included in the CVD event prediction cohort. A total of 2,513 also had CCTA in the study period. After excluding cases in which the CCTA did not report a CAC score (n = 200), 2,313 patients were included in the CAC estimation cohort (Figure 1).

CVD PREDICTION.

The CVD prediction cohort included 27,990 patients with a median age of 63.0 years (IQR: 44.0–77.0 years), of whom 51.7% were female (Table 1). Four percent (n = 1,110) experienced CVD events (myocardial infarction, ischemic stroke, heart failure exacerbation) during follow-up (median 533.0 days after CTH; IQR: 213.0–965.0 days). Cardiovascular events included 600 strokes (54.1%), 311 myocardial infarctions (28.0%), and 199 heart failure hospitalizations (17.9%).

The CTH+PREVENT model achieved a C-index of 0.82 (95% CI: 0.78–0.85) compared with 0.75 (95% CI: 0.70–0.79) for PREVENT-CVD variables alone. CTH features alone achieved similar performance to the combined model (Supplemental Table 16).

RISK RECLASSIFICATION.

The CTH+PREVENT model repositioned 99.8% of patients relative to PREVENT-CVD rankings, with 49.7% moving to higher risk and 50.1% to lower risk (Table 2). Patients whose CVD risk predictions were increased by CTH data were younger (median 59.0 vs 66.0 years, P < 0.001) and exhibited more favorable traditional cardiovascular risk profiles, including lower systolic blood pressure (median 130.0 vs 135.0 mm Hg, P < 0.001), better renal function (estimated glomerular filtration rate: median 94.0 vs 84.0 mL/min/1.73 m², P < 0.001), and lower rates of hypertension (15.5% vs 25.1%, P < 0.001), diabetes (6.1% vs 7.9%, P = 0.06), and use of cardiovascular medications including antihypertensives (13.7% vs 19.9%, P < 0.001) and lipid-lowering medications (10.8% vs 16.6%, P < 0.001). On imaging, these patients were more likely to have vascular calcifications (30.2% vs 24.8%, P = 0.001) and infarcts (20.1% vs 5.8%, P < 0.001), but less likely to have atrophy (21.1% vs 29.7%, P < 0.001). Among the 15.7% reclassified between high-/low-risk categories, 74.5% were correctly reclassified. Net Reclassification Index analysis (Supplemental Figure 4) showed superior risk stratification using CTH, particularly in the months immediately following the ED visit. Kaplan-Meier survival analysis (Supplemental Figure 5) demonstrated superior discrimination between high- and low-risk groups after inclusion of CTH data. The distributions of model predictions for stroke and myocardial infarction/heart failure events are similar and therefore reassuring against bias for the prediction of any particular event (Supplemental Figure 6). Results were consistent when the analysis was restricted to patients without stroke as the CT indication (Supplemental Table 17). To quantify the relative information content of CTH vs PREVENT features, we computed variance-based measures of relative explained variation and created visualizations of risk distributions following Harrell’s framework for assessing added predictive value (Supplemental Methods, Supplemental Figures 7 and 8, Supplemental Table 18).

TABLE 2.

Demographics and Clinical Characteristics of Patients by Predicted CVD Risk Score Ranking Changes After Adding CTH to PREVENT Variables

	Rank Up (n = 1,391)	Rank Down (n = 1,402)	P Value
Age, y	59.0 (37.0–74.0)	66.0 (48.0–80.0)	<0.001
Female	792 (56.9)	658 (46.9)	<0.001
Male	597 (42.9)	741 (52.9)	N/A
SBP, mm Hg	130.0 (117.0–145.0)	135.0 (121.0–149.4)	<0.001
Total cholesterol, mg/dL	160.0 (127.0–194.0)	164.0 (130.0–203.5)	0.27
HDL cholesterol, mg/dL	49.0 (38.0–59.0)	50.0 (40.0–64.0)	0.16
eGFR, mL/min/1.73 m2	94.0 (76.0–110.0)	84.0 (63.0–99.0)	<0.001
Obesity	199 (14.3)	240 (17.1)	0.04
Hypertension	216 (15.5)	352 (25.1)	<0.001
Diabetes	85 (6.1)	111 (7.9)	0.06
Antihypertensive medicine	191 (13.7)	279 (19.9)	<0.001
Lipid-lowering medications	150 (10.8)	233 (16.6)	<0.001
Smoking status	73 (8.4)	84 (8.5)	0.39
Vascular calcifications	420 (30.2)	347 (24.8)	0.001
Small vessel disease	568 (40.8)	546 (38.9)	0.31
Atrophy	294 (21.1)	417 (29.7)	<0.001
Infarcts	280 (20.1)	81 (5.8)	<0.001
Abnormality	1,258 (90.4)	1,167 (83.2)	<0.001

Values are median (IQR) or n (%). “Rank Up” indicates patients repositioned to higher risk rankings (higher predicted risk relative to others); “Rank Down” indicates repositioning to lower risk rankings. P values from chi-square tests (categorical variables) and Mann-Whitney U tests (continuous variables). Patients ranked higher with CTH were younger with favorable clinical profiles but more subclinical vascular disease on imaging.

CTH = head computed tomography; N/A = not applicable; PREVENT = Predicting Risk of cardiovascular disease EVENTs; other abbreviations as in Table 1.

CAC ESTIMATION.

The CAC estimation cohort included 2,313 patients with a median age of 65.0 years (IQR: 55.0–74.0), of whom 53.5% were female (Table 1). Calcium scores were distributed as follows: 0 Agatston units (n = 799, 34.5%), 1–10 (n = 187, 8.1%), 11–100 (n = 418, 18.1%), 101–400 (n = 412, 17.8%), and >400 (n = 497, 21.5%). The CTH+PREVENT model achieved a C-index of 0.76 (95% CI: 0.72–0.80) vs 0.75 (95% CI: 0.70–0.79) for PREVENT-CVD alone (Supplemental Table 19). Precision-recall curves at clinically relevant thresholds are provided in Supplemental Figure 9. Performance remained consistent in sensitivity analysis restricted to patients with CTH before CCTA (Supplemental Table 20). To assess robustness to this modeling choice, we performed a sensitivity analysis using log-transformed continuous CAC scores [log(CAC + 1)] (Supplemental Table 21).

RISK RECLASSIFICATION.

The CTH+PREVENT model repositioned 97.8% of patients relative to PREVENT-CVD rankings (Table 3), with 48.5% moving to higher risk and 49.3% to lower risk. Patients reclassified to higher risk were predominantly female (76.8% vs 42.1%, P < 0.001) with higher HDL cholesterol (50.0 vs 42.5 mg/dL, P = 0.02). Net Reclassification Index analysis (Supplemental Figure 4) demonstrated incremental reclassification across CAC thresholds. Patients whose risk rankings increased with CTH were predominantly female (76.8% vs 42.1%, P < 0.001) and demonstrated higher HDL cholesterol levels (50.0 vs 42.5 mg/dL, P = 0.02). Attention map visualization confirmed that the model focused on these intracranial vascular calcifications when making risk predictions (Figure 2).

TABLE 3.

Demographics and Clinical Characteristics of Patients by Predicted CAC Score Ranking Changes After Adding CTH to PREVENT Variables

	Rank Up (n = 112)	Rank Down (n = 114)	P Value
Age, y	65.0 (56.0–75.0)	66.0 (58.0–73.0)	0.91
Female	86 (76.8)	48 (42.1)	<0.001
Male	26 (23.2)	66 (57.9)	N/A
Sex: Unknown	0 (0.0)	0 (0.0)	N/A
SBP, mm Hg	138.0 (127.0–149.6)	134.0 (121.0–147.0)	0.09
Total cholesterol, mg/dL	171.0 (137.5–199.5)	171.0 (130.5–199.5)	0.64
HDL cholesterol, mg/dL	50.0 (42.0–69.5)	42.5 (35.2–62.5)	0.02
eGFR, mL/min/1.73 m2	82.0 (59.0–95.5)	87.0 (70.0–97.0)	0.28
Obesity	25 (22.3)	25 (21.9)	0.94
Hypertension	54 (48.2)	55 (48.2)	1.00
Diabetes	22 (19.6)	18 (15.8)	0.45
Antihypertensive medicine	28 (25.0)	32 (28.1)	0.60
Lipid-lowering medications	22 (19.6)	20 (17.5)	0.69
Smoking status	8 (8.1)	4 (4.1)	0.25
Vascular calcifications	49 (43.8)	49 (43.0)	0.91
Small vessel disease	53 (47.3)	50 (43.9)	0.60
Atrophy	31 (27.7)	23 (20.2)	0.19
Infarcts	19 (17.0)	17 (14.9)	0.67
Abnormality	100 (89.3)	100 (87.7)	0.71

Values are median (IQR) or n (%). “Rank Up” indicates patients repositioned to higher risk rankings; “Rank Down” indicates repositioning to lower risk rankings. P values from chi-square tests (categorical) and Mann-Whitney U tests (continuous). Patients ranked higher with CTH were predominantly female with higher HDL cholesterol.

Abbreviations as in Tables 1 and 2.

FIGURE 2. Attention Map Visualization.

Representative CTH slices from 2 patients with high coronary artery calcium scores (predicted 4, ground truth 3 for top row; predicted 3, ground truth 3 for bottom row). Top row: Attention maps highlight calcifications in the right greater than left distal internal carotid arteries in a single patient. Bottom row: Attention maps highlight calcification in the distal right internal carotid artery in a second patient. Left panels show the pretrained DINOv2 model with diffuse attention across multiple brain regions. Right panels show the same model after fine-tuning on CAC estimation, demonstrating substantially more focused attention on areas of intracranial arterial calcification (visible as hyperdense regions in the brain window CT images). Attention maps are overlaid on brain window CT images using a heat map (red = high attention, black = low attention). Abbreviations as in Figure 1.

TEMPORAL VALIDATION.

To assess model generalizability across time, we performed temporal validation using data from August 2020 to December 2023 for training and January to August 2024 for testing. For CVD prediction (Supplemental Table 22), temporal validation demonstrated preserved discrimination, with the CTH model achieving a C-index of 0.77 (95% CI: 0.71–0.82), outperforming PREVENT-CVD (C-index 0.71; 95% CI: 0.64–0.77) by 0.06 (0.03–0.09). For CAC estimation, the CTH+PREVENT model maintained robust performance with a C-index of 0.77 (95% CI: 0.73–0.80), comparable to the 0.76 achieved in standard random splits (Supplemental Table 23). Performance remained consistent across all CAC thresholds. These findings confirm that CTH-derived features maintain predictive value in temporally distinct cohorts but highlight the importance of adequate follow-up duration for CVD event prediction.

DISCUSSION

CTH is the most common CT examination in the ED, and can be useful beyond the narrow clinical question (such as the presence of acute hemorrhage or stroke) that prompts its acquisition.¹¹ We evaluated the potential of routinely acquired CTH to improve cardiovascular risk screening, finding that CTH features improve the estimation of coronary artery calcification and major cardiovascular events compared with standard clinical risk factors. In particular, CTH resulted in the correct reclassification of younger patients with low-risk clinical features but high-risk imaging profiles.

Notably, CTH features provided substantial incremental predictive information for CVD events beyond PREVENT-CVD variables. An important consideration is that the PREVENT-CVD model was developed for 10-year risk prediction, whereas our median follow-up was 533 days (approximately 1.5 years). This shorter follow-up likely explains the PREVENT-CVD model’s lower C-index (0.75) compared with its published performance. Traditional clinical risk factors accumulate their effects over years to decades, whereas CTH may capture more proximate markers of subclinical vascular disease—such as silent infarcts and calcifications—that reflect current disease burden and may be more proximally associated with near-term events. Model calibration demonstrated good performance at earlier time points (3 months: Brier = 0.01, ICI = 0.05; 6 months: Brier = 0.02, ICI = 0.09), although calibration decreased at longer follow-up (2 years: Brier = 0.11, ICI = 0.26). This implies that meaningful CVD risk assessment can be performed in the absence of data such as lipid panels. Neuroimaging captures direct evidence of subclinical end-organ damage and atherosclerotic burden that complements upstream demographic and laboratory risk markers. Recent validations report only moderate discrimination from PREVENT-CVD,²³ reinforcing the need for complementary modalities such as CTH. CTH-based risk assessment could be of particular benefit to high-risk patients without access to routine preventive care.²⁴ Indeed, patients without a primary care provider are disproportionately imaged with CT in the ED;²⁵ using these studies to opportunistically estimate CVD risk can increase the value and impact of their ED care. By embedding an automated risk-scoring model into routine clinical workflows, clinicians could identify high-risk patients in real time without the additional radiation or cost of CCTA. If validated prospectively, CTH-based risk reclassification could inform post-ED clinical pathways, serving as an opportunistic signal for downstream cardiovascular evaluation.

Patients reclassified to higher cardiovascular risk despite favorable traditional risk profiles may benefit from automated notifications, expedited referral to primary care or cardiology, or further outpatient risk assessment such as dedicated CAC scoring. For patients without a primary care provider, overrepresented in ED imaging, CTH-based screening could serve as a critical touchpoint for initiating preventive care that might otherwise be missed.

Several mechanisms may explain the relationship between CTH and cardiovascular risk. First, intracranial and skull base calcifications may co-occur with systemic vascular calcification, serving as a proxy for broader atherosclerotic burden.²⁶ Second, age-related structural brain changes, such as white matter disease or cerebral atrophy, may reflect cumulative exposure to vascular risk factors.^27–29 Finally, the quantity¹¹ and standardization of CTH imaging in acute care settings provides a uniquely high-yield substrate for AI models to extract yetundescribed patterns that correlate with cardiovascular outcomes.

Our exploratory analyses provide insight into which patient populations may benefit the most from the addition of CTH to cardiovascular risk stratification. In the CVD prediction setting, patients reclassified to higher risk were younger and exhibited fewer traditional risk factors, such as hypertension and diabetes. Notably, this group had a higher prevalence of infarcts and vascular calcifications on CTH despite the exclusion of individuals with prior stroke diagnosis, suggesting the presence of silent brain infarcts, which are known to reflect CVD risk.^5,30,31 In CAC estimation, patients reclassified to higher risk were predominantly female, with higher HDL cholesterol. These findings suggest that CTH can complement traditional clinical risk factors by revealing evidence of subclinical end-organ damage.

STUDY STRENGTH AND LIMITATIONS.

The strength of our study lies in the integration of robust imaging and longitudinal outcome data across 2 cardiovascular endpoints for a large and diverse ED population.^16,17 Nonetheless, our study has several limitations. First, our cohort consisted of ED patients from a single, tertiary academic center, whose characteristics and risks may not generalize to the general population. Further validation in multicenter and community-based cohorts is needed to ensure the model’s generalizability. Second, our lookback period for prevalent CVD begins in August 2020, although prior diagnoses persist in the electronic health record through problem lists and past medical history. However, care fragmentation may result in systematic underascertainment of CVD history and risk factors, as 63% of ED patients have a documented primary care provider within our health system, but patients may receive care across multiple systems not captured in our electronic health record. This fragmentation particularly affects laboratory data (80% missing lipid panels) and may result in some patients with prevalent CVD from outside systems being misclassified. Importantly, this limitation affects traditional clinical risk models and imagingbased approaches equally, and the consistency of CTH performance across multiple imputation approaches and complete case analysis (Supplemental Tables 6 to 9) suggests our findings are robust to different missingness patterns. Third, the study relies on International Classification of Diseases codes in the electronic health record to define CVD outcomes, which may introduce misclassification bias. Some PREVENT variables were unavailable in our dataset and were replaced by alternatives (for example, body mass index was replaced by obesity status), potentially leading to an underestimation of the PREVENT-CVD’s predictive performance. Fourth, our follow-up period, although sufficient to demonstrate predictive capability, is shorter than the 10-year timeframe used in established CVD risk assessment tools. Longer-term validation studies will be essential to fully establish the clinical utility of CTH-based risk stratification and to enable direct comparison with standard 10-year risk prediction models. Fifth, patient reclassification to a higher risk category after analysis of CTH may reduce specificity, or identify different patient phenotypes, than conventional risk factors alone. Future work must assess whether patients reclassified to higher risk based on CTH in fact benefit from preventive interventions. Conversely, although CTH features are predictive of CAC scores, the correlation is imperfect, and the degree of reassurance provided by a low-risk CTH requires further study. Finally, this is a retrospective study and cannot evaluate the impact of AI-based risk reclassification on clinical decision-making or outcomes.

CONCLUSIONS

We find that routinely performed CTH studies contain cardiovascular risk information that complements traditional clinical risk factors and can significantly enhance cardiovascular risk stratification by identifying subclinical vascular disease not captured by conventional risk models. This approach offers a promising avenue for opportunistic cardiovascular risk screening at scale, enabling earlier identification and treatment of high-risk patients not captured by traditional clinical models.

PERSPECTIVES.

COMPETENCY IN MEDICAL KNOWLEDGE:

CTH imaging contains latent cardiovascular risk information that exceeds traditional clinical risk factors, with deep learning models identifying subclinical vascular disease in younger patients with otherwise favorable risk profiles.

COMPETENCY IN PATIENT CARE:

Integration of automated cardiovascular risk assessment from routine CTH into ED workflows could enable opportunistic screening and preventive interventions for millions of patients annually without additional cost, radiation exposure, or workflow disruption.

TRANSLATIONAL OUTLOOK 1:

Prospective studies are needed to determine whether patients reclassified to higher cardiovascular risk based on CTH features benefit from early preventive interventions such as statins, antihypertensives, or lifestyle modifications.

TRANSLATIONAL OUTLOOK 2:

Further investigation should elucidate the specific CTH imaging features (intracranial vascular calcifications, silent brain infarcts, white matter disease) that drive cardiovascular risk prediction to better understand underlying pathophysiological mechanisms linking brain and cardiovascular health.

ABBREVIATIONS AND ACRONYMS

AHA

American Heart Association

AI

artificial intelligence

CAC

coronary artery calcium

CCTA

coronary computed tomography angiography

C-index

concordance index

CT

computed tomography

CTH

head computed tomography

CVD

cardiovascular disease

ED

emergency department

HDL

high-density lipoprotein

ICI

Integrated Calibration Index

APPENDIX

For supplemental methods, tables, and figures, please see the online version of this paper.

CODE AVAILABILITY

Code for model training and evaluation is publicly available at https://github.com/dkimlab/headct-cvd-prediction.