Prediction Model to Optimize Long-Term Antithrombotic Therapy Using Covert Vascular Brain Injury and Clinical Features

Abstract

BACKGROUND:

Defining the risk of developing major bleeding, especially intracranial hemorrhage (ICH), or ischemic stroke (IS) in patients receiving antithrombotic therapy is crucial. Existing risk prediction tools would inadequately assess the net clinical benefit of antithrombotic therapy. We aimed to develop novel risk scores incorporating covert vascular brain injury to personalize the risk assessment of major bleeding, ICH, and IS in patients receiving antithrombotic therapy.

METHODS:

The prospective, multicenter, observational study (BAT2 [Bleeding With Antithrombotic Therapy Study-2]) enrolled patients receiving oral antiplatelets or anticoagulants from 52 hospitals across Japan between 2016 and 2019. Multimodal brain magnetic resonance imaging was performed at baseline under prespecified conditions to determine cerebral small vessel disease (white matter hyperintensity, cerebral microbleed, lacune, enlarged perivascular space, and cortical superficial siderosis), nonlacunar infarct, and intracranial artery disease with central reading. Risk scores, collectively termed the BAT2 scores, were developed separately to evaluate the comparative risks of (1) major bleeding, (2) ICH, and (3) IS based on covariates from Cox proportional hazards models and clinical relevance. Model performance was assessed with the Harrell C-index and calibration slope adjusted for optimism via bootstrapping.

RESULTS:

Of 5378 patients enrolled, 5250 were analyzed (mean age, 71±11 years, 33% women); 93 experienced major bleeding, including 55 had ICH, and 197 had IS during a median follow-up of 2.0 years. Predictors for bleeding included age, underweight, renal impairment, hypertension, cerebral microbleed, lacune, and antithrombotic treatment type. Predictors for ICH further included deep white matter hyperintensity but not renal impairment. For IS, predictors included age, renal impairment, diabetes, atrial fibrillation, lacune, cerebral microbleed, nonlacunar infarct, and intracranial artery disease. Prediction performance showed optimism-adjusted C-index and calibration slope of 0.69 (95% CI, 0.64–0.74) and 0.82 (95% CI, 0.62–1.06) for bleeding, 0.75 (95% CI, 0.67–0.80) and 0.80 (95% CI, 0.56–1.02) for ICH, and 0.64 (95% CI, 0.60–0.68) and 0.92 (95% CI, 0.73–1.18) for IS.

CONCLUSIONS:

The BAT2 scores may help optimize the balance between risks and benefits of antithrombotic therapy.

REGISTRATION:

URL: https://www.clinicaltrials.gov; Unique identifier: NCT02889653. URL: https://www.umin.ac.jp/ctr; Unique identifier: UMIN000023669.

Long-term antithrombotic therapy with oral anticoagulants or antiplatelet agents (APs) increases bleeding risk. Notably, antithrombotic therapy–related intracranial hemorrhage (ICH) has increased in recent decades partly due to an increasingly aging population.^1,2 Therefore, refined risk stratification to personalize antithrombotic therapy is crucial for shared clinical decision-making. Most bleeding prediction models have primarily been developed and validated for patients using vitamin K antagonists or AP and, more recently, for those on direct oral anticoagulants (DOACs).^3–7 Existing bleeding scores, focusing primarily on clinical features, underperform in elderly individuals, where conventional risk factors show weaker associations with outcomes due to multiple comorbidities, competing nonvascular mortality risks, and overlapping risk factors for ischemic and bleeding events,⁸ and are accordingly only weakly encouraged in clinical guidelines.^9,10 Because ischemic events are more frequent than bleeding events,^11–14 identifying an appropriate prediction model is critical to reducing bleeding risks while maintaining therapeutic efficacy.

Covert vascular brain injuries, such as cerebral small vessel disease (SVD), are incidentally detected by magnetic resonance imaging (MRI) during standard clinical care.^15,16 SVD commonly refers to deep perforator arteriopathy or cerebral amyloid angiopathy, both of which are primary pathological substrates for future stroke.¹⁷ In addition, SVD burden has been associated with an increased risk of extracranial hemorrhage.^18,19 Neuroimaging features of SVD encompass diverse phenotypes, each reflecting chronic vascular vulnerabilities and inherently distinct underlying mechanisms.^20,21 Apart from SVD, intracranial artery disease (ICAD) independently contributes to ischemic stroke (IS) risk.²² Consequently, these underlying vascular brain injuries collectively predispose individuals to an increased risk of cerebrovascular events.

The Microbleeds International Collaborative Network (MICON) group proposed a model of antithrombotic-related ICH or IS using clinical variables and cerebral microbleed (CMB) on MRI, offering improved discrimination.²³ Meanwhile, no studies have investigated the full spectrum of SVD and ICAD within the same individuals. It remains unclear whether predictive models incorporating this imaging can be developed for risk stratification.

To enhance risk prediction, we aimed to develop models to estimate the net benefit of antithrombotic therapy.

Methods

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Study Design and Participants

BAT2 (Bleeding With Antithrombotic Therapy Study-2) was an investigator-initiated, prospective, multicenter, observational study involving 52 hospitals across Japan through the Network for Clinical Stroke Trials.²⁴ BAT2 was designed to determine the incidence and characteristics of bleeding complications in patients receiving oral antithrombotic therapy. The study was registered with https://www.clinicaltrials.gov (Unique identifier: NCT02889653) and the University Hospital Medical Information Network clinical trial registry in Japan (Unique identifier: UMIN000023669). The overall protocol has been published elsewhere.^25,26 Patients with cerebrovascular or cardiovascular diseases, either symptomatic or asymptomatic, who initiated or continued oral antithrombotic therapy, were enrolled in stroke clinics and hospitals between 2016 and 2019. All procedures were approved by the ethics committees of the participating sites, and written informed consent was obtained from patients or their family members before enrollment. The selection choice and dosage of antithrombotic agents were determined by the responsible treating investigator. Antithrombotic therapies at baseline included single-antiplatelet therapy (aspirin, clopidogrel, or cilostazol), dual-antiplatelet therapy, warfarin, and DOAC (dabigatran, rivaroxaban, apixaban, and edoxaban). Brain MRI was mandatory for all participants at registration, and contraindications for MRI were exclusion criteria. The study protocol permitted the inclusion of participants who underwent brain MRI within 14 days after registration. This study adheres to the STROBE reporting guidelines (Strengthening the Reporting of Observational Studies in Epidemiology) for cohort studies.

Data Acquisition

Baseline clinical information and blood test results were collected. Brain MRI, using either a 3T or 1.5T magnetic field, was obtained parallel to the anterior commissure-posterior commissure line or the orbitomeatal line. MRI was performed at registration (within 90 days before or up to 14 days after registration), with standardized imaging protocols across all facilities. MRI sequences, including T1-weighted, T2-weighted, fluid-attenuated inversion recovery, and T2*-weighted imaging, were conducted along with 3-dimensional time-of-flight magnetic resonance angiography. Details regarding the acquisition and interpretation of MRI are described elsewhere.²⁶ All MRI examinations were interpreted by a central diagnostic committee, blinded to clinical information, for CMB, white matter hyperintensity, basal ganglia perivascular space (BG-PVS), lacune, and nonlacunar infarct, according to the criteria of Standards for Reporting Vascular Changes in Neuroimaging.^20,26 The study protocol defined the following criteria for data collection: the number of CMB was categorized as 0, 1, 2, 3, 4, or ≥5; the grade of periventricular hyperintensity as 0, 1, 2, or 3; the grade of deep and subcortical white matter hyperintensity (DSWMH) as 0, 1, 2, or 3; the number of BG-PVS as 0, 1 to 10, or ≥11; and the number of lacunes as 0, 1, 2, 3, 4, or ≥5.²⁶ Confluent white matter hyperintensity was classified as positive when the grade of periventricular hyperintensity was 3 or the DSWMH was 2 or 3.²⁷ Cortical superficial siderosis was identified as curvilinear hypointense lesions along the cortical surface.^26,28 ICAD was assessed using magnetic resonance angiography and classified into 4 categories: normal to mild (signal reduction <50%), moderate (≥50%), severe (focal signal loss at the stenotic lesion with distal signal present), and occlusion in M1 and M2 segments of the middle cerebral artery, A1 and A2 segments of the anterior cerebral artery, intracranial vertebral artery, basilar artery, and P1 and P2 segments of the posterior cerebral artery.^26,29 Interrater reliability of MRI interpretation by the central diagnostic radiology committee, expressed as median kappa coefficients, is given as follows: 0.87 (0.72–0.97) for deep CMB; 0.86 (0.74–0.96) for lobar CMB; 0.68 (0.57–0.86) for periventricular hyperintensity; 0.75 (0.63–0.81) for DSWMH; 0.61 (0.52–0.81) for BG-PVS; and 0.75 (0.65–0.98) for lacune.²⁶ Intrarater reliability for these findings showed median kappa coefficients ranging from 0.66 to 0.94. Detailed kappa values and concordance rates for MRI findings have been reported in our previous study.²⁶

Outcomes

The outcomes were major bleeding, as defined by the International Society on Thrombosis and Haemostasis,³⁰ and IS. Major bleeding included ICH and extracranial major bleeding, with ICH further analyzed as a separate outcome, encompassing intracerebral, subdural, subarachnoid, and extradural hemorrhages. Patients were followed up at 6 (±2), 12 (±3), 18 (±3), and 24 (±4) months after enrollment, either in-person at the clinic or via telephone interviews conducted with the patients or their caregivers by the investigator or trained study coordinators at the participating sites. Patients were instructed to report any events promptly to the investigator in charge to ensure timely data collection. Outcome events were determined by the investigator at each participating site based on neurological findings and relevant imaging features. For diagnoses made at external institutions, clinical records were obtained to verify the details of the event. The investigator adjudicated the determination of the occurrence and classification of events.

Statistical Analysis

The 2-year event risk was determined in the study design. We censored patients at the last available follow-up or at the time of the outcome event. If a patient experienced a first episode of major bleeding (including ICH) or IS, follow-up was censored at that time. Predictor variables, prespecified in the protocol papers,^25,26 were selected based on prior evidence of association with bleeding risk, IS, or clinical rationale for their ability to predict events. To develop the risk model, CMBs were initially classified into deep and lobar categories according to prespecified measurements; however, for ease of interpretation, deep and lobar CMBs were subsequently combined in the risk model. Categorizing CMB or lacunes into 0, 1 to 4, and ≥5 stratified risks while maintaining sufficient sample sizes in each category and ensuring clinical relevance and interpretability. Age was categorized into 2 groups: <65 and ≥65 years for bleeding risk and <85 and ≥85 years for IS, based on restricted cubic spline analysis. The association of each outcome with potential predictor variables was assessed using univariate Cox regression analysis, retaining all variables with a significance threshold of P<0.25. A backward stepwise selection approach was then used with the remaining candidate variables to determine the final model based on the Akaike information criterion. Potential variables with clinical relevance but low event counts were retained in the final model. Each predictor’s regression coefficient was tripled and rounded to the nearest integer. We used a scoring system to present the final model comprising 3 subtypes, termed the BAT2 score (BAT2-bleeding, BAT2-ICH, or BAT2-IS). Model discrimination was evaluated using the Harrell C-index. Calibration was evaluated at 2 years by the slope of calibration plots. The C-index and calibration slope were penalized for optimism using bootstrapped resamples (1000 iterations). Models were fitted to the entire cohort, and their performance was evaluated using internal-external 5-fold cross-validation. The cohort was split into 5 groups based on the study site, and in each fold, the model was evaluated in one group while being trained on the remaining 4. This evaluation process was repeated across all 5 group combinations and iterated 20 times to assess the variability and robustness of model performance.^31,32 We evaluated the observed and predicted risks of each outcome across 3 risk groups categorized as low (0–3 for bleeding/ICH and 0–1 for IS), medium (4–6 for bleeding/ICH and 2–3 for IS), and high risk (≥7 for bleeding/ICH and ≥4 for IS) using Kaplan-Meier estimates, as well as 9 risk groups, with higher-scoring individuals combined into one score category.

We compared the predictive performance of the BAT2 scores with existing risk scores for bleeding and IS in this cohort. For bleeding risk, we evaluated HAS-BLED, ATRIA, ORBIT, DOAC, MICON-ICH, and S₂TOP-BLEED scores.^3–7,23 For IS, we assessed CHADS₂, CHA₂DS₂-VASc and MICON-IS scores.^23,33,34 Statistical differences were determined using the DeLong method and pairwise comparisons of the C-index. Although random missingness was detected across variables, the percentage of missing data remained below 2%; therefore, no imputation techniques were applied in developing risk models. As a sensitivity analysis to examine potential bias from missingness, we repeated the model development procedure after imputing missing values using multiple imputations with chained equations (20 imputations). Results are reported in accordance with the TRIPOD guideline (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis).³⁵ Statistical analyses were performed using Stata (version 18.0) and R (version 4.2.0).

Results

Among the 5378 patients registered, 11 were excluded due to MRI contraindications, 17 had no MRI data acquisition for reasons other than contraindication, 24 had incomplete baseline clinical data, 2 were duplicated registrations, 73 lacked follow-up data, and 1 had no information on baseline antithrombotic therapy (Figure 1). Thus, 5250 patients (mean age, 71±11 years; median age, 73 [interquartile range, 66–79] years; 1736 [33.0%] women; 5247 [99.7%] Asians) were included in the analyses. Of these, 4727 (90.0%) had a history of IS or transient ischemic attack, with a median of 73 days (interquartile range: 15–1446) from onset to registration, and were taking antithrombotic agents for secondary stroke prevention. The remaining 523 (10.0%) received antithrombotic therapy for primary stroke prevention or secondary prevention of cardiovascular diseases. Baseline antithrombotic regimens included 871 patients (16.6%) on DOAC alone, 433 (8.2%) on warfarin alone, 3134 (59.7%) on single-antiplatelet therapy alone, 551 (10.5%) on dual-antiplatelet therapy alone, and 263 (5%) on oral anticoagulant and AP.

Figure 1.

Patient flow diagram. BAT2 indicates Bleeding With Antithrombotic Therapy Study-2; and MRI, magnetic resonance imaging.

Table 1 shows the baseline characteristics of the entire patient population. The 5250 patients provided 9933 patient-years of follow-up (median, 2.00 [interquartile range, 1.80–2.03] years). A total of 93 major bleedings, including 55 ICH and 38 extracranial major bleedings (24 gastrointestinal bleedings), and 197 IS were observed. Of the 55 ICH cases, intracerebral hemorrhage accounted for 45.0% (24 cases), followed by subdural hematoma (20 cases) and subarachnoid hemorrhage (9 cases). Of the 197 IS cases, 97 cardioembolic strokes, 42 noncardioembolic strokes, and 57 other or undetermined strokes were included.

Table 1.

Baseline Characteristics

The hazard ratios of each predictor variable are shown in Table S1. Predictors retained in the final model for major bleeding, ICH, and IS, along with the resulting integer risk scores, are presented in Table 2. After stepwise selection, the following predictors of major bleeding and ICH were identified: age ≥65 years, body mass index <18.5 kg/m², CMB (1–4 and ≥5), lacune (≥5), and antithrombotic treatment type (single-antiplatelet therapy, dual-antiplatelet therapy, DOAC, warfarin, and oral anticoagulant+AP). Hypertension and estimated glomerular filtration rate <30 mg/dL were retained in the bleeding model based on clinical judgment. In the ICH model, DSWMH was further incorporated via stepwise selection, whereas the estimated glomerular filtration rate <30 mg/dL did not show an association and was excluded. For the IS model, the predictors included age ≥85, diabetes, atrial fibrillation, estimated glomerular filtration rate <30 mg/dL, CMB (1–4 and ≥5), lacune (1–4 and ≥5), and severe ICAD after stepwise selection.

Table 2.

Multivariable Hazard Ratios for Risk of Major Bleeding, Intracranial Hemorrhage, and Ischemic Stroke From the Final Model

The final model for each outcome showed moderate discrimination, with the optimism-adjusted C-index of 0.69 (95% CI, 0.64–0.74) for major bleeding, 0.75 (95% CI, 0.67–0.80) for ICH, and 0.64 (95% CI, 0.60–0.68) for IS. The calibration slope was 0.82 (95% CI, 0.62–1.06) for major bleeding, 0.80 (95% CI, 0.56–1.02) for ICH, and 0.92 (95% CI, 0.73–1.18) for IS (Table 3). After internal-external validation, model performance remained consistent for major bleeding (C-index, 0.67 [SD, 0.09]; calibration slope, 0.77 [SD, 0.46]); ICH (0.72 [SD, 0.06] and 0.75 [SD, 0.25]); and IS (0.64 [SD, 0.05] and 0.95 [SD, 0.39]; Table 3).

Table 3.

Discrimination, Calibration, and Validation of the BAT2 Score for Each Outcome

The calibration results for each outcome across 3 or 9 clinical risk groups demonstrated that predicted events were aligned with observed risks (Figures 2 and 3). The BAT2-bleeding, BAT2-ICH, and BAT2-IS scores each outperformed existing bleeding or ischemic risk scores (all P_difference<0.001). All existing risk scores had a C-index <0.6, except for the MICON-ICH score (0.67; Table 4).

Figure 2.

Kaplan-Meier plots and risk tables for major bleeding, intracranial hemorrhage, and ischemic stroke. Shaded areas represent 95% CI. Risk scores ranged from 0 to 16 for major bleeding, 0 to 20 for intracranial hemorrhage (ICH), and 0 to 13 for ischemic stroke (IS). The risk categories were defined as low (0–3 for bleeding/ICH and 0–1 for IS), moderate (4–6 for bleeding/ICH and 2–3 for IS), and high (≥7 for bleeding/ICH and ≥4 for IS). Predictors for major bleeding included age (≥65 years), underweight (body mass index [BMI] <18.5 kg/m²), renal impairment (estimated glomerular filtration rate [eGFR] <30 mL/min), hypertension, type of antithrombotic therapy (single-antiplatelet therapy [SAPT], dual-antiplatelet therapy [DAPT], direct oral anticoagulant [DOAC] alone, warfarin alone, and oral anticoagulants+antiplatelet), cerebral microbleed (CMB; number; 1–4 and ≥5), and lacune (number; ≥5). Predictors for intracranial hemorrhage included age (≥65 years), underweight (BMI <18.5 kg/m²), hypertension, type of antithrombotic therapy (SAPT, DAPT, DOAC alone, WF alone, and oral anticoagulants+antiplatelet), CMB (number; 1–4 and ≥5), lacune (number; ≥5), and deep and subcortical white matter hyperintensity (grade ≥2). Predictors for ischemic stroke included age (≥85 years), renal impairment (eGFR <30 mL/min), diabetes, atrial fibrillation, CMB (number; 1–4 and ≥5), lacune (number; 1–4 and ≥5), nonlacunar infarct, and severe intracranial arterial stenosis (focal signal loss at the stenotic lesion with distal signal present or occlusion). A, Major bleeding; (B) intracranial hemorrhage; (C) ischemic stroke.

Figure 3.

Major bleeding, intracranial hemorrhage, and ischemic stroke model calibration by risk scores. Predicted vs observed risk of each event by risk score. Predictors for major bleeding included age (≥65 years), underweight (body mass index [BMI] <18.5 kg/m²), renal impairment (estimated glomerular filtration rate [eGFR] <30 mL/min), hypertension, type of antithrombotic therapy (single-antiplatelet therapy [SAPT], dual-antiplatelet therapy [DAPT], direct oral anticoagulant [DOAC] alone, warfarin alone, and oral anticoagulants+antiplatelet), cerebral microbleed (CMB; number; 1–4 and ≥5), and lacune (number; ≥5). Predictors for intracranial hemorrhage included age (≥65 years), underweight (BMI <18.5 kg/m²), hypertension, type of antithrombotic therapy (SAPT, DAPT, DOAC alone, WF alone, and oral anticoagulants+antiplatelet), CMB (number; 1–4 and ≥5), lacune (number; ≥5), and deep and subcortical white matter hyperintensity (grade ≥2). Predictors for ischemic stroke included age (≥85 years), renal impairment (eGFR <30 mL/min), diabetes, atrial fibrillation, CMB (number; 1–4 and ≥5), lacune (number; 1–4 and ≥5), nonlacunar infarct, and severe intracranial arterial stenosis (focal signal loss at the stenotic lesion with distal signal present or occlusion). A, Major bleeding; (B) intracranial hemorrhage; (C) ischemic stroke.

Table 4.

Comparison of the Discrimination of the BAT2 Score With Existing Risk Scores

In sensitivity analyses, the multiple imputation method identified the same predictor candidates, and model performance remained consistent.

Discussion

We developed the different BAT2 scores to stratify each outcome among patients receiving various antithrombotic agents. The BAT2 score accounts for the higher cumulative risk associated with different combinations of antithrombotic therapy and vascular burden, including SVD and ICAD. The BAT2-bleeding and BAT2-IS models discriminated moderately, whereas the BAT2-ICH demonstrated relatively higher discrimination in bootstrap internal-external validation. Calibration was excellent across all models, with particularly good agreement between predicted and observed outcomes in the BAT2-IS model. Each BAT2 score significantly improved performance compared with existing risk scores.

The BAT2 bleeding score excludes variables known to be associated with a bleeding risk due to their extremely low prevalence in this cohort (eg, bleeding diatheses, liver disease, or nonsteroidal anti-inflammatory use) or potential collinearity with other predictors, which may have masked their significance. Considerable overlap exists between clinical variables in existing bleeding risk tools and the BAT2 bleeding score, including age, hypertension, low body mass index, renal impairment, and antithrombotic agents. Previous risk scores generally had a C-index below 0.60, with the exception of MICON-ICH, suggesting that achieving high discriminative ability using only conventional clinical features remains challenging. In addition, the differences in risk scores likely reflect the heterogeneity of the studied patient populations. Predicting major bleeding is inherently difficult, as systemic bleeding involves diverse mechanisms and risk factor profiles. Our data suggest that incorporating intermediate markers, such as covert vascular brain injury, may improve predictive performance. This is explained by the fact that SVD and ICAD mirror the extent of vascular injury, both of which are linked to future clinical events and demonstrate potential clinical utility.²¹

The BAT2-ICH score demonstrated a C-index of 0.75, closely aligning with the discrimination of the MICON-ICH score (0.73) observed in the MICON cohort.²³ However, the predictive value for ICH was underestimated when validating the MICON-ICH in the BAT2 cohort (C-index, 0.67). Notably, the BAT2-ICH score improved discrimination by further incorporating severe nonhemorrhagic SVD (lacune ≥5 and DSWMH ≥2). Lipohyalinosis, characterized by wall thickening and luminal narrowing with fibrinoid necrosis, is believed to be a common mechanism by which hypertension induces SVD. In later stages, this process can lead to occlusion or rupture of microvessels, presenting as lacune or CMB.¹⁵ A meta-analysis found that lacune and white matter hyperintensity each confer a 3-fold increased risk of ICH.³⁶ Lacunar infarction has been included as a component in existing bleeding risk scores, such as HAS-BLED and Intracranial-B₂LEED₃S.^34,37 Thus, nonhemorrhagic SVD likely holds prognostic value for ICH, leading to more accurate risk prediction. In contrast, neither BG-PVS nor cortical superficial siderosis was incorporated into the BAT2 score. This exclusion aligns with previous findings that BG-PVS may have a less pronounced clinical impact than other SVDs, likely due to differing underlying pathologies, including brain arterial aging.³⁸ Despite cortical superficial siderosis being a hemorrhagic hallmark of the cerebral amyloid angiopathy–related phenotype, its exclusion may be attributed to the patient characteristics and its low prevalence in this study although the observed prevalence was similar to that reported in previous pooled studies.^39,40

Although several predictors, such as advanced age, CMB ≥1, lacune ≥5, and renal impairment, were commonly associated with both bleeding and IS, outcome-specific predictors for IS were also identified. For IS prediction, diabetes and a history of IS are common predictors included in existing IS tools and the BAT2-IS. In addition, the presence of lacune ≥1 or severe ICAD was identified as being specifically predictive of IS risk. These findings highlight the importance of evaluating both shared and outcome-specific risk factors when considering antithrombotic strategies. In our analysis, the discriminatory performance of the BAT2-IS was modest (C-index, 0.64). Prediction of IS is likely challenging due to the diverse etiologies with distinct underlying mechanisms, which may contribute to the difficulty in developing a highly predictive risk model. In contrast, calibration performance was consistently excellent (calibration slope, 0.92), suggesting that the BAT2-IS may have potential clinical utility in predicting actual risk.

Although the cutoff value in the current clinical consensus has yet to be established, annual incidence rates of major bleeding ≥3%, ICH ≥1%, or IS ≥2% (in the context of anticoagulation therapy) are generally indicative of high-risk groups.^8,41,42 Based on these definitions, the predicted 2-year cumulative incidence is estimated at 5.99% for major bleeding, 1.99% for ICH, and 3.96% for IS. These rates correspond to BAT2-bleeding ≥9/16, BAT2-ICH ≥9/20, and BAT2-IS ≥3/13, provisionally categorizing them as high-risk groups. Using the BAT2 score to quantify bleeding and IS risk enables physicians to balance these risks and consider treatment modifications such as avoiding bleeding risk factors, de-escalating antithrombotic potency, implementing strict blood pressure control, or exploring nonpharmacological options for emboli prevention.

The strength of our study lies in the large sample size and multicenter recruitment in a real-world setting, reflecting actual clinical experiences and prescribing patterns. Second, we used standardized multimodal MRI sequences, rated for SVD and ICAD, using validated scales by a central diagnostic radiology committee, ensuring high reliability in neuroimaging interpretation. Third, event information was collected not only for major bleeding or ICH but also for IS, enabling the development of individual risk models for antithrombotic-related outcomes. Fourth, internal validation using bootstrapping and internal-external cross-validation were conducted following the TRIPOD guidelines.³⁵

Our study has limitations. First, bleeding or ischemic risk may vary between individual AP or DOAC agents and their exposure times, potentially affecting predictive performance. Patients may have switched or discontinued treatment during the follow-up period. Our study was based on observational data and limited to evaluating the effects of different antithrombotic medications at baseline, precluding causal interpretation. Second, the assessment of PVS was limited to the BG, and we did not assess the centrum semiovale, despite PVS in the centrum semiovale being associated with cerebral amyloid angiopathy.³⁸ Third, we were unable to fully assess the performance of the BAT2-ICH score for intracerebral hemorrhage due to low event counts. Fourth, although we compared the BAT2 score to existing risk scores, comparison against an independent external cohort would clarify the relative performance of each BAT2 score. Fifth, almost all participants were Asian, which might limit the generalizability of the results to other ethnicities. Our scores were also derived from a cohort with a relatively short follow-up period, a low proportion of women, and no central adjudication of outcome events or independent external validation; these factors may further limit their applicability for external validity. Outcome adjudication was not blinded to clinical information, which may have introduced ascertainment bias. Sixth, model development relied on a stepwise backward selection. Although effective for variable selection, it may introduce limitations such as biased coefficient estimates and misleading CIs. Finally, we did not develop separate prediction models for antiplatelet or anticoagulant therapies. However, our models could be applied clinically by estimating the predicted risks for antithrombotic-related events for different antithrombotic treatment options, enabling a comparative assessment to inform therapeutic decision-making.

With the incorporation of covert vascular brain injury, each BAT2 score demonstrates adequate discrimination for risk stratification of antithrombotic-related events in patients receiving various antithrombotic therapies. As bleeding and IS have inherently opposing treatment implications with respect to antithrombotic therapy, separate evaluation of risk scores for each outcome is essential to support appropriate clinical decision-making. By separately quantifying and concurrently evaluating these risks, it may support a more balanced risk-benefit assessment and aid in tailoring antithrombotic treatment strategies aimed at achieving net clinical benefit. The clinical utility of the BAT2 scores warrants validation in prospectively designed studies.

Article Information

Sources of Funding

BAT2 (Bleeding With Antithrombotic Therapy Study-2) was organized by a central coordinating center located at the National Cerebral and Cardiovascular Center, with funding support from the Japan Agency for Medical Research and Development (grants JP18ek0210055, JP24lk0221171, and JP24lk0221186) and the Japan Society for the Promotion of Science (grant JP23K27522).

Disclosures

All of the following conflicts are outside the submitted work. Dr Koga reports grants from Boston Scientific Corporation, Nippon Boehringer Ingelheim Co Ltd, and Daiichi Sankyo Company Ltd; compensation from Bristol Myers Squibb, AstraZeneca, Bayer, Daiichi Sankyo Company, Otsuka Pharmaceutical Co Ltd, and Pfizer for other services; and compensation from Janssen Pharmaceuticals for consultant services. Dr Yakushiji reports grants from Bristol Myers Squibb, Daiichi Sankyo, Bayer, and Pfizer Japan. Dr Fujimoto reports grants from Daiichi Sankyo Company Ltd, Bristol Myers Squibb, Bayer, Daiichi Sankyo Company Ltd, Nippon Boehringer Ingelheim Co Ltd, and Pfizer and compensation from Bayer, Daiichi Sankyo Company Ltd, Nippon Boehringer Ingelheim Co Ltd, Pfizer, and Bristol Myers Squibb for other services. Dr Yagita reports compensation for the lecture fee for other services. Dr Shinichi Yoshimura reports grants from Medtronic, Bristol Myers Squibb, Medico’s Hirata, Otsuka Pharmaceutical Co Ltd, Terumo, Eisai Co Ltd, Stryker, and Daiichi Sankyo Company Ltd and compensation from Bayer, Johnson & Johnson Health Care Systems, Inc, Kaneka Medics, and Idorsia for other services. Dr Hirano reports compensation from Bayer for other services and compensation from Daiichi Sankyo for other services. Dr Toyoda reports compensation from Daiichi Sankyo and Janssen Pharmaceuticals for other services. The other authors report no conflicts.

Supplemental Material

Table S1