A Genomic Risk Score Identifies Individuals at High Risk for Intracerebral Hemorrhage

Abstract

Background:

Intracerebral hemorrhage (ICH) has an estimated heritability of 29%. We developed a genomic risk score for ICH and determined its predictive power in comparison to standard clinical risk factors.

Methods:

We combined genome-wide association data from individuals of European ancestry for ICH and related traits in a meta-genomic risk score (metaGRS) (2.6 million variants). We tested associations with ICH and its predictive performance in addition to clinical risk factors in a held-out validation dataset (842 cases and 796 controls). We tested associations with risk of incident ICH in the population-based UK Biobank cohort (486,784 individuals, 1,526 events, median follow-up 11.3 years).

Results:

One SD increment in the metaGRS was significantly associated with 31% higher odds for ICH (95%CI: 1.16-1.48) in age-, sex- and clinical risk factor-adjusted models. The metaGRS identified individuals with almost 5-fold higher odds for ICH in the top score percentile (OR: 4.83, 95%CI: 1.56-21.2). Predictive models for ICH incorporating the metaGRS in addition to clinical predictors showed superior performance compared to clinical risk factors alone (c-index: 0.695 vs 0.686). The metaGRS showed similar associations for lobar and non-lobar ICH, independent of the known APOE risk locus for lobar ICH. In the UK Biobank, the metaGRS was associated with higher risk of incident ICH (HR: 1.15, 95%CI: 1.09-1.21). The associations were significant within both a relatively high-risk population of antithrombotic medications users, as well as among a relatively low-risk population with a good control of vascular risk factors and no use of anticoagulants.

Conclusions:

We developed and validated a genomic risk score that predicts lifetime risk of ICH beyond established clinical risk factors among individuals of European ancestry. Whether implementation of the score in risk prognostication models for high-risk populations, such as patients under antithrombotic treatment, could improve clinical decision making should be explored in future studies.

Graphical Abstract

INTRODUCTION

Intracerebral hemorrhage (ICH) is the most devastating type of stroke. Although it accounts for 10-20% of all acute cerebrovascular events, it is responsible for almost 50% of stroke-related morbidity and mortality.¹ Given the lack of effective treatments and the devastating outcome of ICH, primary and secondary prevention is critical. As with many complex human diseases, ICH risk is comprised of both environmental and genetic factors. Early genome-wide association studies (GWAS) of ICH estimated a heritability of 29% and revealed a polygenic architecture,^2,3 whereas large-scale GWAS for established risk factors for ICH, such as hypertension and smoking, have identified hundreds of associated genomic loci that cumulatively explain a large proportion of the variance of these traits.^4-9

Beyond defining heritability and providing insight into biological mechanisms, GWAS findings have begun to show promise for disease risk prediction. Genomic risk scores (GRS), biomarkers representing the aggregated effect of many genetic variants on a given trait, have been proposed as powerful tools for identifying individuals at high risk for complex traits, with a predictive performance at times comparable to that of rare monogenic mutations.¹⁰ A GRS for ICH could have clinical utility in decision making algorithms, such as for complementing risk-benefit calculation tools in patients prescribed antithrombotic medications. Despite acknowledged limitations, pertaining mainly to translation of the accuracy of GRS findings from the cohort- to the individual-based level, as well as sex- and ancestry-specific predictive differences,^11,12 if constructed according to best practices,^11-14 an ICH GRS could have important implications for patient selection for clinical trials. However, as opposed to other vascular diseases,¹⁰ efforts to construct a GRS for ICH lag behind, likely because of the lower statistical power in available datasets due to the relative rarity of the disease.

Recent advances in analytical approaches offer more efficient alternative methods for GRS construction that allow to utilize data from multiple GWAS in order to overcome power limitations of genomic datasets of rare diseases.^15,16 Based on the assumption that the majority of genetic variants exert their effects on a given disease by affecting intermediate traits, constructing a GRS based on genomic data of traits in the causal pathway of the disease of interest can improve both power and prediction.¹⁵ In this study, we sought to investigate whether combining genetic liability for possible ICH risk factors and traits reflecting pathologies underlying ICH into a meta-Genomic Risk Score (metaGRS) could improve our ability to predict ICH events among individuals of European ancestry. To further explore the potential utility of such a genomic score for clinical risk prediction, we also assessed whether the metaGRS improves ICH risk prediction beyond established clinical risk factors.

METHODS

Data availability statement

The data that support the findings of this study will be available from the corresponding author upon reasonable request. MetaGRS single nucleotide polymorphism (SNP)-specific weights will be made publicly available at The Polygenic Score (PGS) Catalog.

Study design and participating studies

As our primary data source, we used genotype and phenotype data from 1,861 ICH cases and 1,722 ICH-free controls from three independent GWAS datasets: the North American (USA) multi-center Genetics of Cerebral Hemorrhage on Anticoagulation (GOCHA) study, a prospectively collected case-control study of European ancestry subjects aged > 55 years with primary ICH¹⁷; the European member sites contributing ICH cases and controls to the International Stroke Genetics Consortium (EUR/ISGC); and the Genetic and Environmental Risk Factors for Hemorrhagic Stroke (GERFHS) study, a prospectively collected case-control study of subjects > 18 years of age with spontaneous ICH in the Greater Cincinnati region¹⁸. GOCHA and EUR/ISGC were the training datasets, whereas GERFHS was our primary validation dataset. Furthermore, we performed an external validation of the derived score for incident ICH events in the UK Biobank (UKBB) cohort, over a median follow-up of 11.3 years among 486,623 individuals aged 40-69 years at recruitment without a prior history of ICH.¹⁹ All studies were approved by the Institutional Review Board or Ethics Committee at each participating institution. We followed the GRS reporting guidelines as outlined in the Polygenic Risk Score Reporting Standards developed by the Clinical Genome Resource (ClinGen) Complex Disease Working Group and the PGS Catalog¹⁴. Details on ascertainment of cases and controls and on imputation of the GWAS datasets are provided in Supplemental Material.

Trait-specific GRS and ICH metaGRS construction

We used GOCHA (436 ICH cases and 405 controls) and EUR/ISGC (577 ICH cases and 523 controls) as training datasets to develop GRS for 21 traits associated with ICH risk. We leveraged publicly available GWAS summary-level data from international consortia, as detailed in Table S1.^4,20-32 For all traits, we used data from European-only populations and excluded duplicate and ambiguous AT/GC SNPs and SNPs with MAF≤1%. No GOCHA or EUR/ISGC cases or controls were included in the selected studies. For each trait, we generated a range of candidate trait-specific GRS and selected the best-performing ones based on the highest area under the receiving operating characteristics curve of a logistic regression model for ICH. Following a standardized meta-analytic approach, we constructed an ICH metaGRS creating a weighted-average of the trait-specific GRS. We used GERFHS as our primary validation dataset. Details are provided in Supplemental Methods, Tables S2-S23.

ICH metaGRS performance and clinical evaluation in GERFHS and UKBB

We explored the metaGRS performance in the validation dataset and its comparison with clinical predictors. The primary metaGRS and alternative versions were entered as linear predictors in logistic regression models for ICH in the primary validation dataset. We performed complementary analyses to investigate the association between the metaGRS and ICH. We evaluated the potential clinical utility of the metaGRS exploring its associations with ICH independently of clinical risk factors in GERFHS and performed external validation in a population-based setting in the UKBB (Supplemental Methods).

RESULTS

Construction of a metaGRS for ICH in the training dataset

A schematic of our study design is provided in Figure 1. Following quality checks, we developed 21 optimized GRS for ICH-associated traits on the basis of associations with ICH in GOCHA and EUR/ISGC. The numbers of variants included in these GRS ranged from 213 to 1,148,192 (Table S23). Detailed results are provided in the Supplemental Results, Tables S24-S25, and Figures S1-S4.

Figure 1. Study design.

(A) Individual genomic risk scores (GRS) were derived for intracerebral hemorrhage (ICH)-related traits from publicly available summary statistics. (B) The GRS were optimized in the combined training dataset of GOCHA and EUR/ISGC. (C) MetaGRS parameters were determined on the basis of association with ICH in the training datasets. (D) The metaGRS was compiled in the GERFHS validation dataset and associations with ICH were explored using logistic regression. (E) The metaGRS was externally validated in the general population-based UKBB cohort using Cox proportional hazards regression analyses.

Associations between the metaGRS and ICH in the validation dataset

We explored associations between the derived metaGRS and ICH risk in GERFHS. Characteristics of ICH cases and ICH-free controls in GERFHS are presented in Table S26. After adjusting for age, sex, and two PCs, one SD increase in the metaGRS was associated with 45% higher odds of ICH (OR 1.45; 95% CI: 1.30-1.63; p=6.2x10⁻¹¹). Patients in higher thresholds of the metaGRS distribution were at progressively higher risk for ICH (Figure 2). Notably, patients in the top 2.5% and 1% had substantially increased odds of ICH, respectively, compared to the rest of sample. We observed an expected gradual change in ICH odds when moving to either the higher or lower ends of the metaGRS distribution from the middle decile (Figure S5). Modeling risk of ICH relative to the bottom 10% of the metaGRS loading did not reveal any significant non-linear effects (Figure S6).

Figure 2. Odds for intracerebral hemorrhage (ICH) across the metaGRS distribution.

Depicted is the metaGRS distribution (centered around a mean of 0 and a standard deviation of 1) in the GERFHS validation dataset and the odds ratios (OR) for ICH per percentile group, relative the rest of the sample, as derived from logistic regression models adjusted for age, sex and the first 2 principal components of the population structure.

In analyses comparing ICH odds between the primary and the alternative metaGRS, we overall found that the primary metaGRS achieved the best performance (Figure S7). However, these metaGRS with penalized and adjusted weights still demonstrated significant associations with ICH risk, albeit with weaker effect size estimates and lower predictive performances (Figure S7, Table S27).

Predictive performance of metaGRS for ICH in comparison with clinical risk factors

We next explored the performance of the metaGRS with established clinical risk factors for predicting ICH in GERFHS. History of ischemic stroke, hypertension, diabetes, high cholesterol, heavy alcohol use, anticoagulant use, and education less than high school were most strongly associated with ICH risk (Table S28). After adjusting for these clinical risk factors, the metaGRS continued to be independently associated with ICH risk (OR = 1.31 per one standard deviation of the metaGRS; 95% CI 1.16 – 1.48 p < 0.0001). Adding the metaGRS to a model including the clinical risk factors significantly improved the model fit (AIC of clinical predictors = 1989.65, AIC of clinical predictors + metaGRS = 1972.86, LRT = 4x10⁻⁵).

The metaGRS showed comparable predictive performance to hypertension, and higher than the remaining clinical risk factors apart from education (Figure 3). Importantly, the c-index of a model including the entire set of clinical risk factors (C: 0.686, 95%CI: 0.663-0.718) increased after including the metaGRS in the model (C: 0.695, 95%CI: 0.673 – 0.727, Figure 3). Similar results were observed for the alternative metaGRS versions (Figure S8, Table S29). Stratified analyses showed significant associations of the metaGRS with both lobar and non-lobar ICH (Tables S30-S34). When exploring in the same model the metaGRS and APOE genotype, a known risk locus for ICH, we found that both were independently associated with the odds of ICH (Table S35).

Figure 3. Performance of clinical risk factors, metaGRS, and their combination for predicting odds for intracerebral hemorrhage (ICH).

Depicted are the c-indices derived from logistic regression models including age, sex, 2 principal components of population structure (baseline model, dotted line) and additionally in successive models the reported clinical variables (history of ischemic stroke, hypertension, diabetes, high cholesterol, heavy alcohol use, anticoagulant medication use, and education) and the metaGRS in the validation GERFHS dataset. Percentile confidence intervals of the c-indices were calculated after bootstrapping over 1000 iterations.

Validation of the metaGRS in the UKBB population

As a final step, we explored associations between the metaGRS and incident ICH risk in a general population sample. In the prospective population-based UKBB cohort, a total of 486,623 participants without a history of ICH, were followed-up for a median of 11.3 years (IQR: 10.6-11.1 years) (Table S36). We again found the metaGRS to be associated with a higher risk of incident ICH (HR per SD increment: 1.15, 95%CI: 1.09-1.21, p=7x10⁻⁷) after adjusting for age, sex, the first 10 PCs of population structure, kinship, genotyping chip, and genetic ancestry. Accounting for death as a competing risk with the subdistribution hazard approach did not substantially alter the results (HR per SD increment of metaGRS: 1.14, 95%CI: 1.08-1.20, p=4x10⁻⁶). Figure 4A presents the Kaplan-Meier curves with age as the time variable for individuals at upper, median, and lower quantiles of the metaGRS. Within both high-risk individuals using antithrombotic medications at baseline, as well as low-risk individuals with well-controlled vascular risk factors at baseline and no antithrombotic medications, the metaGRS retained association with incident ICH risk (Figures 4B, C). We again found the metaGRS to be associated with a higher risk of incident ICH (HR per SD increment: 1.15, 95%CI: 1.09-1.21, p=7x10⁻⁷, concordance: 0.705, 95%CI: 0.693-0.717) in a model adjusted for age, sex, the first 10 PCs of population structure, kinship, genotyping chip, and genetic ancestry.

Figure 4. MetaGRS and cumulative risk of incident intracerebral hemorrhage (ICH) in the population-based UK Biobank sample.

The results are derived from (A) entire UKBB population, (B) users of antithrombotic medications at baseline, and (C) low-risk individuals with conventional vascular risk factors under control and no use of anticoagulant medications after excluding prevalent cases of ICH. Depicted are the Kaplan-Meier curves for different metaGRS quantiles, as well as the hazard ratios (HR) per standard deviation (SD) increment in metaGRS, as derived from Cox proportional hazards regression models adjusted for baseline age, sex, the first 10 principal components of the population structure, genetic ancestry, genotyping chip, and kinship.

DISCUSSION

We developed a genomic risk score for ICH in a training dataset of 1,013 cases and 928 controls based on GWAS data for 21 ICH-related traits. We found the derived metaGRS to be significantly associated with the odds of ICH in an independent validation dataset of 842 ICH cases and 796 ICH-free controls. The metaGRS was independent of traditional clinical risk factors of ICH and improved model performance in prediction of ICH. Furthermore, the score was significantly associated with incident ICH risk in a population-based cohort study of 480,000 individuals followed-up for a median of 11 years (1,500 incident ICH events). Our results provide important insights into genomic prediction for ICH and could have implications for clinical practice.

First, the metaGRS identified individuals at very high risk for ICH. For example, individuals at the top percentile had almost 5-fold increased odds for ICH, as compared to the rest of the population. While it remains to be clarified how these individuals would benefit from potential primary preventive interventions, this information could be useful both for screening for hypertension, the main clinical risk factor for ICH, and early initiation of antihypertensive treatment, as well as for decision making when considering initiation of antiplatelet or anticoagulation treatments that might increase ICH risk. These risk stratification strategies based on genomic information are increasingly important as millions of persons in the US and around the world have been genotyped by direct-to-consumer genotyping companies.

Second, the metaGRS improved risk discrimination for ICH when compared to classical clinical predictors. Specifically, it was associated with ICH risk independently of vascular risk factors and was found to have a predictive value superior to all predictors except for education. The predictive power of the metaGRS was comparable to that of hypertension, the clinical risk factor for ICH that explains the most variance in the trait.³³ These findings support the incorporation of genetic information into clinical tools aiming to quantify ICH risk within specific patient subgroups. A post-hoc analysis of trial data showed that among patients with atrial fibrillation and a CHA2-DS2-VASc score of 2, a high genomic risk score for ischemic stroke led to an absolute ischemic stroke risk equivalent to those with a higher score.³⁴ Whether integration of a genomic risk score for ICH in such analyses could lead to a more precise assessment of the risk-benefit ratio for specific patients remains to be determined. Along these lines, the several clinical trials currently evaluating anticoagulation as a secondary prevention strategy after ICH constitute a unique opportunity for genomic-based risk-stratification, as a portion of them have built-in biobanks that are collecting DNA samples.

Third, despite its rarity, we found the metaGRS to be significantly associated with prospective ICH risk in the general population. As expected, the strength of association was attenuated in the relatively healthy population of UKBB, compared to the disease-focused case-control setting where the metaGRS was developed. The metaGRS was associated with a higher risk of ICH even among individuals with evidence-based control of relevant risk factors, who were not actively smoking, had blood pressure of 140/90 mmHg or less, no evidence of diabetes, normal BMI, and who reported no use of anticoagulants. While such analyses are restricted by lack of power, our results suggest that for individuals with a high genetic risk, the recommended treatment targets for modifiable risk factors might not be sufficient for primary ICH prevention. The importance of this observation lies on the fact that the genomic information is available long before risk factors are present and could thus be used for earlier risk stratification in otherwise low-risk individuals. Concomitantly, the metaGRS was also associated with a higher risk of ICH even among a high-risk group of individuals using antithrombotic medications, indicating its potential utility among a relevant group of patients for whom bedside calculation of ICH risk might be particularly relevant to clinical care.

Our study has limitations. First, the sample sizes of the available genetic datasets for ICH are limited, as compared to other clinical endpoints. This introduces uncertainty to the association estimates between the genetic variants and ICH risk, which were used to construct the metaGRS, and thus impacts negatively on its predictive performance. Indeed, when compared with metaGRS for other traits that have been developed in larger datasets, such as coronary artery disease and ischemic stroke,^15,16 the association with incident ICH events is weaker. Additionally, while previous studies exist exploring the genetic architecture of ICH and cerebral small vessel disease that could potentially be used for construction and comparison of our genomic risk score, these were either derived from the same population sample as the one we use in the current study, are ICH GWAS in ancestrally diverse populations, or derived from candidate gene studies^2,3,35. As such, utilizing the above studies to construct a GRS would result in significant over-estimation of association estimates¹³ and limit the GRS accuracy partly due to linkage disequilibrium and allele frequency differences across populations³⁶, as well as by including a limited number of non-replicated loci³⁷. Second, ICH is a phenotypically heterogeneous disease, with the most common etiologies being hypertensive small vessel disease (typically in non-lobar locations) and cerebral amyloid angiopathy (typically in lobar locations). To maximize the power of our approach, we have pooled cases, which could have negatively impacted the predictive performance for specific ICH etiologies. While our score was predictive for both non-lobar and lobar ICHs, developing etiology-specific scores might be of more relevance for specific clinical scenarios. Furthermore, in the training and the validation datasets, we did not account for dose of risk factors, due to either suboptimal phenotyping in the datasets, or to avoid including a larger number of clinical variables than already included, which would likely contribute to model instability, respectively. Third, while the metaGRS showed significant associations with risk of incident ICH in the UKBB, we could not explore its effects in concert with other clinical predictors, because the metaGRS was generated using datasets including data from the UKBB. Therefore, independent validation either of a score trained in an entirely UKBB-independent dataset or of the described metaGRS in another external cohort would be necessary. Fourth, the metaGRS was constructed solely based on data from individuals of European genetic ancestry, and may thus not be applicable for individuals of other ancestries. Larger multi-ethnic GWAS studies of ICH currently underway will facilitate the generation of ancestry-specific GWAS datasets. Furthermore, because our metaGRS is derived from genetic susceptibility to common vascular risk factors, it is likely to carry associations with other cardiovascular traits as well, despite being trained on ICH. This is a fundamental limitation of the technique, and our metaGRS is therefore not intended as a tool to identify risk of ICH at the exclusion of other endpoints, such as coronary artery disease. However, this expected pleiotropic effect of the metaGRS allows for pursuing follow-up research, focused on composite outcomes that share common biological underpinnings. Last, overarching limitations and challenges still exist on the generation and validation of GRS across disease states, which apply to our metaGRS as well. Some of them include the possible differences in sex-specific predictive performances, the translation of GRS estimates from the cohort- to the individual-specific level which has been suggested to introduce additional variability, as well as the heterogeneity of the different methods for GRS construction which could ultimately hinder clinical application.^11,12 Towards that end, efforts are currently underway to standardize and delineate procedures surrounding GRS construction and reporting, such that prediction models incorporating GRS-based estimations can be leveraged in a consistent and reproducible manner.¹⁴

In conclusion, our study represents the first comprehensive attempt to develop and validate a genomic risk score for ICH. Our results demonstrate that the incorporation of genomic information in clinical prediction models for ICH could enhance predictive performance. As such, it lays the groundwork for future analyses in larger genetic datasets for ICH to optimally combine genomic information to maximize predictive benefit. Exploration of the performance of genomic risk scores for ICH in clinical trials of patients receiving antithrombotic medications could offer useful insights in risk prediction of ICH in this high-risk population with potential relevance for clinical decision making.

Non-standard Abbreviations and Acronyms

ICH

intracerebral hemorrhage

GWAS

genome-wide association study

GRS

genomic risk score

SNP

single nucleotide polymorphism

GOCHA

Genetics of Cerebral Hemorrhage on Anticoagulation

GERFHS

Genetic and Environmental Risk Factors for Hemorrhagic Stroke

EUR/ISGC

European member sites contributing to the International Stroke Genetics Consortium

UKBB

UK Biobank

QC

quality control

PC

principal component

HRC

Haplotype Reference Consortium

MAF

minor allele frequency

LD

linkage disequilibrium

AUC

area under the receiving-operating characteristics curve

CAD

coronary artery disease

AIC

Akaike Information Criterion

HES

hospital episode statistics

SBP

systolic blood pressure

DBP

diastolic blood pressure

PP

pulse pressure

WMH

white matter hyperintensities

CKD

chronic kidney disease

eGFR

estimated glomerular filtration rate

UACR

urine albumin-to-creatinine ratio

TC

total cholesterol

TG

triglycerides

LDL

low-density lipoprotein

HDL

high-density lipoprotein

T2D

type 2 diabetes mellitus

HbA1c

hemoglobin A1c

BMI

body mass index

WHR

waist-to-hip ratio

SVS

small vessel stroke

EA

educational attainment