A Risk Prediction Model for Ischemic Stroke in Southern Chinese Population: Impact of Multiple Genetic Variants and Clinical/Lifestyle Factors

With a global prevalence of 7.7 million, ischemic stroke (IS) is one of the leading causes of death and disability worldwide. In China, IS alone contributed to 69.6% of stroke events and accounted for 37.1% of the mortality/disability rate^[1]. IS is a complex disease that is known to be associated with various genetic variants and clinical and lifestyle risk factors^[2]. Genome‐wide association studies (GWAS) provided evidence for the occurrence of more than 160 IS‐associated single nucleotide polymorphisms (SNPs). In terms of clinical risk factors, patients with a history of chronic diseases like hypertension, dyslipidemia, and diabetes mellitus, display a higher risk of developing IS. Among the various lifestyle components, leading a sedentary lifestyle, smoking, and having an unhealthy diet are known to be associated with the risk of IS. Thus, integration of various genetic, clinical, and lifestyle variables might prove to be highly beneficial in the prediction and prevention of IS at the individual level^[3].

The present study aimed to construct a risk prediction model for IS by simultaneous incorporation of both genetic variants and clinical/lifestyle indicators. The study involved a prospective cohort of the Southern Chinese population. It is expected that the proposed model could be validated externally, and the obtained significant features would assist in the interpretation of IS risk in the Chinese population, with a particular focus on IS risk prediction at the individual level.

The study was conducted on subjects belonging to four community health service centers in the Ningbo City of Zhejiang Province. Initially, a total of 2,349 participants aged ≥ 40 years without any history of IS were recruited from April–July 2013. Detailed inclusion and exclusion criteria followed in the present study are described in Supplementary Figure S1 (available in www.besjournal.com). The clinical and physical parameters included in the present study were strictly defined. The information regarding the demographic and lifestyle characteristics of the subjects was collected using a standard questionnaire. The salient features of the questionnaire are demonstrated in Supplementary Table S1 (available in www.besjournal.com). After a follow up period of three years, the updated records for each participant containing information regarding the occurrence of any IS incidence were obtained from the electronic health record database. During this 3-year follow up, the individuals that were first diagnosed with heart failure, atrial fibrillation, or myocardial infarction were excluded from the study as such diagnoses might lead to significant changes in their lifestyles, which might further act as confounding factors in the present study. Consequently, 236 participants were excluded, and a total of 2,113 individuals were included in the study.

Figure S1.  Study design.

Initially, the target SNPs were identified and selected from IS-related genetic studies using common databases. A standard SNP selection process was implemented, which was previously established by Li et al.^[4]. For genotyping, a total of 102 SNPs were selected (Supplementary Table S2 available in www.besjournal.com). For blood sample collection, the participants were subjected to overnight fasting, and samples were drawn by venipuncture, collected in vials containing anticoagulant EDTA, and preserved at −80 ℃. DNA was extracted using Tiangen Blood Genomic DNA extraction kits. For genotyping, the polymerase chain reaction (PCR)/ligase detection reaction (LDR) were adopted. The PCR reactions contained 1 µL genomic DNA, 1.5 µL 10× PCR buffer, 1.5 µL MgCl₂, 0.3 µL dNTPs, 0.15 µL each primer, and 0.2 µL Taq DNA polymerase in a total volume of 15 µL, and were performed in an ABI Prism 7000 Sequence Detection System with an initial melting at 94 °C for 3 min, 35 cycles of denaturation at 94 °C for 15 s, annealing at 55 °C for 15 s, extension at 72 °C for 30 s, and final extension at 72 °C for 3 min. Each ligation reaction included 3 µL PCR product, 1 µL 10× Taq DNA ligase buffer, 5 U Taq DNA ligase, and 0.01 µL each discriminating probe in a total volume of 10 µL, and was carried out in 30 cycles at 94 °C for 30 s and 56 °C for 3 min. Re‐sequencing results for 10% of the samples showed that the concordance rates were > 95% for all target SNPs.

In the present study, elastic net regression was adopted for the modeling process^[5]. In particular, this method introduces ℓ2‐norm and ℓ1‐norm penalties into the regularization term to deal with high correlation variables and estimates a series of coefficients $ {\widehat{\beta }}_{Elastic} $ as per the following Equation:

Here, $ {y}_{i}\in \left\{\mathrm{0,1}\right\} $ denotes the response variable and ${\pi }_{i}=p({y}_{i}=1\left|{x}_{i})\right.=\dfrac{\mathit{exp}\left({x}_{i}^{T}\beta \right)}{1+\mathit{exp}\left({x}_{i}^{T}\beta \right)},i=\mathrm{ }\mathrm{1,2},\cdots ,n.$ The tuning parameters λ₁ and λ₂ determine the regularized logistic regression solution and coefficient estimates. In this study, 102 SNPs and 27 clinical/lifestyle covariates were used in the construction of IS risk prediction model. Individuals’ final risk scores were obtained, and they were further classified into three IS risk categories (high/intermediate/low). In particular, positive predictive values (PPVs), sensitivity, and specificity were calculated. The discriminative ability of the constructed models was measured in terms of the area under curves (AUC). In terms of the given captured predictors, the impact of various risk profiles on IS was evaluated, and R software, version 3.6.1, was used for analysis. Further, for the identification of KEGG/Reactome pathways and gene ontology (GO) function interpretations, the enrichment analysis tool g:Profiler was adopted^[6].

Among 2,113 participants recruited in this study, 3.17% were newly diagnosed with IS by August 2016. Supplementary Table S3 (available in www.besjournal.com) summarizes the baseline features, which were statistically compared between cases and controls. Interestingly, age, height, SBP, clinical history of hypertension, and dietary habits involving the consumption of egg, red meat, chicken, and fish showed significant differences between cases and controls, with P < 0.05.

At the initial genetic‐based modeling stage, elastic net regression was applied on 102 SNPs selected from common databases. Further, the performance of all candidate models was evaluated using 10-fold cross‐validation, and the model with the highest fitted AUC value was recognized as the best model. As shown in Figure 1, the derived/resulting genetic model captured 15 SNPs and was characterized by a fitted AUC of 0.691 (95% CI: 0.627–0.755). Following this, the model‐driven genetic risk scores were calculated for each individual and utilized for the construction of the full model. In the next stage, 27 clinical/lifestyle covariates, including three demographic features, four anthropometric parameters, six clinical measurements, clinical history of two diseases, and 12 lifestyle variables, were used along with the model‐driven genetic risk scores for further modeling. Finally, a complete model for IS risk prediction was generated, with a fitted AUC of 0.846 (95% CI: 0.803–0.89). This model identified four parameters, including age, model‐driven genetic risk score, SBP, and fish intake, as important predictors of IS risk (Figure 1 and Supplementary Table S4 available in www.besjournal.com).

Figure 1.  The ROC Curves of the genetic-based model and the full IS risk prediction model

The resulting full model was further used to calculate the IS-risk scores for each participant. Subsequently, the participants were classified into three risk categories (Supplementary Table S5 available in www.besjournal.com). Eventually, 68.34% of 2,113 participants corresponded to the low‐risk group, with only 0.76% of these participants developing IS during a 3-year follow up period. In comparison to this, the intermediate and high‐risk groups included 27.64% and 4.02% of the participants. Importantly, 5.82% and 25.88% of the subjects belonging to intermediate and high‐risk groups, respectively, developed IS within three years. Interestingly, the risk of developing IS was found to be 34-times higher in participants belonging to the high‐risk category as compared with those categorized into the low‐risk group. These results further highlighted that the generated model displayed a good discriminatory ability to identify patients with a high risk of IS.

Further, the results of the univariate analysis revealed that the four recognized features were independently associated with IS (Supplementary Table S6 available in www.besjournal.com). In particular, individuals aged ≥ 60 years (OR: 2.39, 95% CI: 1.39–4.95), having elevated SBP (regression coefficient: 11.30, 95% CI: 6.66–15.93), or those with increased genetic risk (regression coefficient: 0.71, 95% CI: 0.51–0.91) displayed inflated IS risk. On the contrary, the dietary intake of fish reduced the risk of IS (OR: 0.70, 95% CI: 0.57–0.86). In order to investigate the relative risk of IS, individuals were further classified into various risk profiles. For three captured features of age, SBP, and fish intake, the individuals were defined as “healthy” for clinical/lifestyle exposures if the individual were aged < 60 years, had < 140 mmHg SBP, and consumed fish at least once a week, whereas the subjects were categorized as “intermediate” healthy ones if they fulfilled only two of the three criteria defined for “healthy” exposures. In cases where the subjects fulfilled either one or none of the aforementioned criteria, these were categorized as “unhealthy”. Additionally, the derived genetic risk scores were used to generate three genetic risk strata, wherein the top third were treated as high genetic risk and the bottom third as low risk. As shown in Table 1, the relative risk of IS gradually increased as individuals’ clinical/lifestyle exposures changed from healthy to unhealthy status and genetic susceptibility changed from low to high. Thus, the subjects with “unhealthy” clinical/lifestyle status and “high” genetic risk displayed the highest relative risk of IS (RR: 15.60, 95% CI: 3.75–64.96) as compared to the reference group. Interestingly, the participants with “low” genetic risk but “unhealthy” clinical/lifestyle status displayed higher IS risk than the individuals with “high” genetic risk but “healthy” non‐genetic status (RR: 5.50, 95% CI: 1.20–25.20). These results highlighted a stronger cumulative impact of the non‐genetic exposures on IS risk than the genetic profiles.

Among the 15 captured SNPs, 6 SNPs or their genes, including rs1800961, rs2954029, rs17321515, rs2575876, rs7493, and rs693, were shown to be directly associated with IS, whereas the remaining 9 recognized SNPs or genes affected IS‐related conditions (Supplementary Table S4). In particular, rs4939883 was reported to be associated with HDL‐c level, rs10889353 contributed to variations in TG levels, and rs4299376, PCSK9 (rs11583680), and SLC12A3 (rs11643718) were LDL‐associated variants. Besides these, rs1229984 was a risk factor associated with alcohol dependence that may increase the risk of IS via alcohol‐induced sympathetic activation, whereas ADCY3 (rs10187348) was a risk factor linked to obesity.

For the identification of KEGG/Reactome pathways and GO functional interpretations, enrichment analysis was performed for 15 captured SNPs via g:GOSt module of g:Profiler tool set (Figure 2)^[6]. For multiple testing adjustments, g:SCS threshold was set to 0.001. Consequently, the captured SNPs were enriched to signaling pathways linked to cholesterol metabolism (KEGG: 04979, adjusted P = 2.471 × 10⁻⁶), fat digestion and absorption (KEGG:04975, P_adjusted = 2.225 × 10⁻⁴), plasma lipoprotein assembly (REAC: R‐HSA‐8963898, P_adjusted = 1.821 × 10⁻⁴), and transport of small molecules (REAC:R‐HSA‐382551, P_adjusted = 4.604 × 10⁻⁴). In terms of the biological domain, the captured SNPs aggregated to cholesterol (GO: 0120020, P_adjusted = 4.505 × 10⁻⁵), sterol (GO: 0120015, P_adjusted = 5.298 × 10⁻⁵), and lipid (GO: 0120013, P_adjusted = 5.686 × 10⁻⁴) transfer activities for molecular function (MF) domain. For the biological process (BP) domain, these SNPs functionally enriched to the lipid/cholesterol/sterol homeostasis (GO: 0055088, 0042632, 0055092, with P_adjusted < 6 × 10⁻⁶). According to human phenotype ontology, the identified SNPs were found to be significantly related to certain diseases, including premature coronary artery atherosclerosis, myocardial steatosis, and cerebral artery atherosclerosis.

Figure 2.  The P-value plot of enriched pathways and functional domains for the 15 captured SNPs

In research settings, when collinear predictors greatly outnumber the available number of samples (P > n), ordinary regression is subjected to overfitting and coefficient instability. Comparatively, the use of the elastic net regression model allows us to control the total number of involved variables using the penalty parameter λ and capture groups of potentially highly correlated variables to build a sparse model that is immune to overfitting^[5]. The robustness of elastic net regression in addressing multicollinearity and overfitting has been previously established^[7]. To verify the same in this study, the data were further divided into construction and validation sets at a ratio of 9:1, and the validation procedure was introduced. The newly generated IS-risk model achieved a fitted AUC of 0.835 on the construction set and a validated AUC of 0.81 on the validation set. These values were slightly lower than the fitted AUC of the original model (0.846). These variations might be attributed to a certain degree of overfitting. Besides this, the occurrence of an insufficient number of cases (n = 60) and samples in the construction set after splitting the data could also have acted as contributing factor, resulting in a less‐comprehensive prediction model that identified only a subset of important risk factors and had reduced power of prediction.

Several previous studies attempted the construction of risk prediction models for both “all‐stroke” and/or “IS‐only”. Similar to the present case, these tools were also developed using common lifestyles, medical conditions, or genetic variants involved in lipid/cholesterol metabolism, statin pathways, or cerebral artery atherosclerosis^{[3, 8, 9]}. These findings highlighted that different subtypes of IS could be induced by similar lifestyle and genetic risk factors^[2]. Recently, an atrial substrate model was proposed by Kamel et al.^[10] illustrating that aging and other common vascular risk factors like unhealthy lifestyles may simultaneously be involved in distinct etiologies underlying different IS subtypes. In particular, these factors could induce an abnormal atrial tissue substrate to cause AF and thromboembolic stroke, and also trigger large‐artery atherosclerosis, ventricular systolic dysfunction, or in situ cerebral small‐vessel occlusion, leading to thrombotic stroke. Nevertheless, different subtypes of IS are still characterized by their unique triggers and pathological mechanisms. Therefore, to reveal the unique etiologies of IS subtypes, it is important and necessary to build specific predictive tools for each subtype of IS.

The present study had certain limitations. The original model derived in this study could not be validated owing to the limited number of cases and sample size. Thus, future studies should apply this model to an independent dataset to verify its accuracy. Additionally, the measurements or classifications of some lifestyle factors used in this study were not standardized, which might further limit the application of this model in external settings.

For fatal diseases that affect populations all across the globe, the development of robust, individualized disease risk assessment tools is the first step toward precision medicine and health care. In the present study, a new tool was generated for IS risk prediction, which involved 15 captured SNPs and three clinical/lifestyle predictors. The results of the study highlighted the suitability of this new tool in IS risk recognition in the Chinese population at the individual level. Additionally, this tool might assist in providing valuable information regarding the implications of various factors in IS etiology.

The authors thank the investigators and participants for their contribution to this study.

The authors declare no conflict of interest.

The study was approved by the Medical Ethics Committee of Hangzhou Normal University (No. 2013020).