Background Despite several known risk factors it is still difficult to foresee who will develop a stroke and who will not. Vascular brain damage, visualised with MRI, reflects how the brain tolerates the effects of vascular risk factors and may therefore be relevant in predicting individual stroke risk.
Objective To examine whether the presence of small vessel disease on brain MRI could improve the prediction of stroke beyond the classic stroke risk factors from the 1991 Framingham Stroke Risk Function.
Methods 1007 community-dwelling elderly people, free of stroke at baseline were included in the study. Small vessel disease—that is, the presence of silent brain infarcts (SBI) and white matter lesions (WML), was scored on MRI scans obtained in 1995–6. 10-Year stroke risk prediction was assessed by the C statistic and by reclassification adding SBI and WML to a risk model including the classic stroke risk factors.
Results During 10-years of follow-up 99 strokes occurred. Individual stroke risk prediction significantly improved from 0.73 (95% CI 0.67 to 0.78) to 0.75 (0.69 to 0.80) in men and from 0.69 (0.64 to 0.75) to 0.77 (0.71 to 0.82) in women after inclusion of SBI and periventricular WML to the stroke risk factors. Reclassification occurred mainly in the intermediate stroke risk group (men 26%; women 61% reclassified).
Conclusions Assessment of small vessel disease with MRI beyond the classic stroke risk factors improved the prediction of subsequent stroke, especially in women with an intermediate stroke risk. These findings support the use of MRI as a possible tool for better identifying people at high risk of stroke.
- Primary prevention
- risk factors
- Alzheimer's disease
- cerebrovascular disease
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.
- Primary prevention
- risk factors
- Alzheimer's disease
- cerebrovascular disease
Population-wide better control of stroke risk factors has reduced the incidence of stroke over recent decades.1 However, typical adherence rates are about 50% for drugs and are much lower for lifestyle prescriptions and other more behaviourally demanding regimens.2 To maximise the value of such strategies, it would be helpful to identify people most at risk of stroke. Individual risk prediction tools can help to identify those individuals and motivate them to lower their risk through risk factor modification.3
The Framingham Stroke Risk Function provided a framework for identification of individuals at high risk and is still the most used risk function for stroke.4 This function is based on the now ‘classic’ stroke risk factors: age, gender, smoking, systolic blood pressure, antihypertensive treatment, diabetes mellitus, atrial fibrillation, left ventricular hypertrophy and coronary heart disease.4 Improvement of this function would lead to better prediction of who will develop stroke and who will not. Several potential new predictors have been added to the Framingham predictors, but they have had no or only moderate success in improving stroke prediction.5–9 Such improvement requires the addition of variables that are strongly associated with stroke risk and highly prevalent in the population, but also have an additive component beyond the known stroke risk factors.10
Population-based MRI studies found strong associations between early (asymptomatic) markers of vascular brain damage and the risk of stroke.11 ,12 These markers may reflect how the end organ tolerates the effects of vascular risk factors and may therefore be relevant in predicting stroke risk. No studies thus far have examined the performance of small vessel disease in predicting stroke.
In the population-based Rotterdam Scan Study we examined whether knowledge of the presence or absence of small vessel disease on brain MRI could improve the prediction of stroke beyond a prediction based on the stroke risk factors from the Framingham Stroke Risk Function.
The Rotterdam Scan Study aims to study causes and consequences of brain changes in the elderly.13 ,14 Baseline examinations of the study were carried out from 1995 to 1996. At that time, we made a random selection of 1904 elderly people aged 60–90 years originating from two population-based cohort studies in strata of age (5 years) and sex.15–17 After exclusion of subjects with contraindications to undergo MRI (such as metal clips, pacemaker or claustrophobia), 1717 people were eligible, of whom 1077 participated and gave written informed consent (>97% Caucasian ethnicity). The study was approved by the medical ethics committee of Erasmus MC University Medical Center.
Cerebral infarcts and white matter lesions (WML) on baseline MRI
All participants underwent brain MRI at baseline. We carried out axial T1-weighted, T2-weighted and proton-density scans on 1.5 Tesla MRI scanners (MR Gyroscan, Philips, Best, The Netherlands; and MR VISION, Siemens, Erlangen, Germany). The slice thickness was 5 or 6 mm with an interslice gap of 20%.
Brain infarcts and WML were scored as described previously.18 In brief, brain infarcts were defined as focal hyperintensities on T2-weighted images that were at least 3 mm in diameter and were scored by a single trained doctor who was unaware of the patient's history of stroke and transient ischaemic attack (TIA).18 Two raters scored periventricular and subcortical WML separately. WML were considered to be present if hyperintensities were visible on proton-density and T2-weighted images, without prominent hypointensities on T1-weighted scans. We considered WML to be periventricular if they were directly adjacent to the ventricle; otherwise we considered them subcortical. We scored periventricular WML semiquantitatively in order to obtain a total periventricular score (range 0–9). A total volume of subcortical WML was approximated on the basis of the number and size of the lesions (volume range 0–29.5 ml).13
We obtained a history of stroke and TIA by self-report and by checking medical records of all participants. Stroke was defined as rapidly developing clinical signs of focal disturbance of cerebral function with no apparent cause other than a vascular origin, with a duration of more than 24 h. Subarachnoid haemorrhages were excluded. A stroke was subclassified as ischaemic or haemorrhagic based on a CT or MRI scan. If we could not retrieve enough information to subclassify a stroke as ischaemic or haemorrhagic, it was called unspecified. TIAs were defined as temporary attacks (commonly 2–15 min, maximum 24 h) with focal symptoms, which are attributable to dysfunction of one arterial territory in the brain. An experienced neurologist subsequently reviewed the medical histories and scans and categorised the infarcts as silent or symptomatic. We defined silent brain infarcts (SBI) as evidence of one or more infarcts on MRI, without a history of (corresponding) stroke or TIA. If a prior stroke or TIA did correspond with a lesion, the latter was defined as a symptomatic infarct. Participants with both symptomatic and silent infarcts were categorised in the symptomatic infarct group. In our analyses, we excluded all participants with a clinical stroke (n=70, with or without symptomatic infarcts on MRI) before the baseline evaluation, leaving 1007 participants for the analyses.
Assessment of Framingham predictors
Cardiovascular risk factors were determined by interview and laboratory and physical examination, as previously described.18 Risk factors included in our analyses were systolic blood pressure, smoking, diabetes mellitus, atrial fibrillation, left ventricular hypertrophy and coronary heart disease. The presence of atrial fibrillation or left ventricular hypertrophy was assessed by automated interpretation of an ECG.19 Coronary heart disease was diagnosed if a participant had a history of myocardial infarction, coronary artery bypass graft or percutaneous transluminal coronary angioplasty that was confirmed by ECG or medical records. Medication use was assessed by interview and registered during house visits.
Follow-up for stroke and mortality
We reviewed the medical records of all participants at the general practitioner's office. Information on vital status was obtained from the municipal health authorities and was available for all participants up to January 2007. About half of the participants were continuously monitored for incident stroke through automated linkage of the study database with files from general practitioners, who are the gatekeepers of the Dutch healthcare system. Nursing home doctors' files, files from general practitioners who did not have an automated linkage system and files from general practitioners of participants who had moved out of the district were also checked regularly (every 3 years). For all reported strokes, we recorded information on signs and symptoms, date of onset, duration and hospital stay. For reported events, additional information was obtained from hospital records. To verify all diagnoses, research doctors discussed information on all potential strokes and TIA with an experienced stroke neurologist. Our strict monitoring procedures made it possible to identify virtually all strokes that occurred during the follow-up period, even in participants who had not been referred to a hospital—for example, people living in nursing homes or participants who had a fatal stroke.
The follow-up time was calculated from the date of the MRI scan until the date of stroke, death or end of follow-up. Follow-up time after 10 years was censored and we used only first-ever strokes. Missing values in covariates (left ventricular hypertrophy (118/1007); coronary heart disease (108/1007); others <1%) were imputed using missing value analysis in SPSS.
We used Cox proportional-hazards analysis to corroborate the previously reported association between the presence of small vessel disease markers on brain MRI scans and the risk of stroke.12 We adjusted for the Framingham stroke risk covariates—that is, age, sex (if applicable), smoking (never/former/current), systolic blood pressure, antihypertensive treatment, systolic blood pressure × antihypertensive treatment, diabetes mellitus, atrial fibrillation, left ventricular hypertrophy and coronary heart disease. The 5-year and 10-year risks of stroke were estimated by the Kaplan–Meier method. Assessment of improvement of global model fit was done using likelihood ratio (LR) tests.
To examine how well small vessel disease discriminates between people who will develop stroke and people who will not during 10 years of follow-up, we computed the C statistic for survival data.20 ,21 We calculated these measures for men and women separately.
First, we assessed the predictive performance of the Framingham predictors over a 10-year follow-up period. Subsequently, we examined whether the predictive accuracy increased when information on SBI (no/yes) and WML (continuous) was added to this prediction model. Both models were most optimally fitted to our data using the aforementioned Framingham stroke risk covariates. We derived estimates from our own data rather than using the individual prediction formulas of the Framingham Stroke Risk Score developed by Wolf et al.4 As shown previously, use of these prediction formulas in our data would probably lead to an overestimation of the predictive effects.22 To further minimise any overoptimism in our model, we assessed the internal validity of the final models using the bootstrap resampling technique, generating 1000 bootstrap samples and using the average optimism to correct the predictive performance of the original models.23
We used the Hosmer–Lemeshow test to evaluate whether there was good alignment between predicted and observed strokes across the entire spread of the data (ie, calibration).
The C statistic cannot assess risk differences as it assesses risk rankings. Risk differences may be clinically important. Therefore we calculated the integrated discrimination improvement that can be seen as the difference in discrimination slopes.20
Guideline recommendations to treat or withhold someone from treatment are increasingly based on a person's individual stroke risk.3 To assess whether our new model more accurately stratifies individuals into higher or lower risk categories than the classic stroke risk factors alone, we created reclassification tables. We classified the participants into three risk categories (<5% or low, 5–15% or intermediate and >15% or high) on the basis of an overall risk of 11.4% and roughly even numbers across the different risk categories. First, people were classified into these three categories on the basis of the initial model containing the Framingham predictors. Second, we reclassified people into these three categories on the basis of our new model. Finally, we compared these categories with the initial model and investigated the accuracy of these reclassifications using the Kaplan–Meier method to calculate the 10-year risk of stroke.
The net reclassification improvement offers incremental value over the reclassification as it distinguishes between individuals correctly reclassified and those incorrectly reclassified. It is defined as the difference in proportions moving up and down among those who will have a stroke versus those who will not have a stroke.20
All analyses were performed using the statistical packages SPSS V.15.0 for Windows (SPSS, Inc) and R version 2.8.1 (R Foundation for Statistical Computing, Vienna, Austria).
Baseline characteristics of the study population, overall and by gender, are shown in table 1. During 10-years of follow-up 99 strokes occurred (12 haemorrhagic, 59 ischaemic, 28 unspecified) and 266 people died without having a stroke, leading to a 10-year risk of stroke of 11.4%. Overall, 10-year follow-up was complete for 96.9% of potential person-years. No significant difference in the 10-year probability of stroke was found between men and women (table 2). People with evidence of small vessel disease had a higher risk of stroke during follow-up than those without, even after adjustment for the Framingham predictors (table 2). When SBI, tertiles of periventricular and subcortical WML were all included in the same model, the associations with stroke risk remained for SBI, but diminished for periventricular and especially subcortical WML (data not shown). Global model fit improved significantly after adding SBI, periventricular and subcortical WML to the classic stroke risk factors (men, LR=12 (p=0.003); women, LR=37 (p<0.001).
The C statistic (after bootstrap correction for optimism) for the Framingham predictors was for men 0.73 (95% CI 0.67 to 0.78) and for women 0.69 (95% CI 0.64 to 0.75). Individual stroke risk prediction improved substantially when SBI, periventricular and subcortical WML were included individually in the two models. The change in C statistic for men was 0.02 and for women 0.08 when we added SBI as well as periventricular WML in a model including the Framingham predictors, indicating a substantial improvement of stroke risk prediction. Addition of subcortical WML to a model containing SBI and periventricular WML did not improve the accuracy (table 3).
Hosmer–Lemeshow tests were non-significant (men p=0.600; women p=0.852), indicating good calibration.
The integrated discrimination improvement was 3.6% (p=0.028) for men and 12.6% (p<0.001) for women, suggesting that the addition of small vessel disease improved the discriminatory property of the model for the prediction of stroke.
The number of people in this cohort who were classified as having risks of <5%, 5–15% or >15% on the basis of the Framingham Stroke Risk Function is shown in tables 4 and 5. As subcortical WML (in addition to SBI and periventricular WML) did not improve the accuracy of stroke prediction, subcortical WML were not included in the analyses for risk stratification. In total, 98 out of 485 men (20%) and 207 out of 522 women (40%) were reclassified on the basis of our new model adding SBI and periventricular WML. This reclassification occurred mainly in individuals in the intermediate stroke risk group. Of the 185 men classified in the intermediate risk category, 19 (10%) men were reclassified to the higher and 30 (16%) men to the lower risk category. Of the 241 women initially classified as having an intermediate risk, 38 (16%) women were reclassified to the higher and 108 (45%) women to the lower risk category (tables 4 and 5). This resulted in an net reclassification improvement of the total sample of 6.3% (p=0.264) for men and 28.3% (p=0.002) for women.
We found that small vessel disease improved the accuracy of the Framingham Stroke Risk Function in 10-year stroke risk prediction. Successful reclassification occurred mainly in women and in risk categories that were initially classified as intermediate.
During recent years, several potential new predictors (eg, carotid intima media thickness, ankle brachial index and C-reactive protein) have been added to the Framingham predictors, but so far they have had no or only moderate success in improving stroke prediction.5–9 We are not aware of other studies that report on the additional value of brain imaging markers to the prediction of stroke. Although some other stroke prediction models have been created,24–26 the Framingham Risk Score is still the most validated prediction rule available for stroke. Further validation of other prediction models that may improve stroke prediction is clearly required. Another important point to consider for future studies is that prediction studies usually investigate the incremental value of only one or a few new markers. It will be highly interesting to investigate the improvement in prediction combining several different emerging risk factors—for example, both imaging and serum markers.
Our finding that the stroke risk prediction model improved after addition of periventricular WML, but not after addition of subcortical WML, may indicate coverage of the disease load by periventricular WML alone, rather than in combination with subcortical WML.27
It is known that SBI, subcortical WML and periventricular WML are associated with higher stroke risk independent of cardiovascular risk factors.11 ,12 However, it is important to emphasise the distinction between risk factors and risk predictors of disease. The presence of an association between a risk factor and a disease does not necessarily imply that this risk factor will be useful for risk prediction or that it will benefit clinical practice. Although several studies have reported on markers of vascular brain damage and the risk of stroke,11 ,12 ,28 this is the first study to report on the additional value of brain imaging markers to the prediction of stroke.
Before this new risk function can be implemented in clinical practice, many steps are required before our results can lead to a change in clinical practice. First, replication of our results in another study population is required and in addition, individual risk scores should be created. Moreover, competing risks with other diseases and mortality should be taken into account and time-varying exposures and non-linear models should be explored. Finally—perhaps most importantly—the cost-effectiveness of MRI screening needs to be investigated. As a result, the clinical and societal implications of our results are not yet clear and require further investigation. However, our results can serve as basis and theoretical proof of concept that stroke risk prediction can be improved beyond the currently best available tool.
There was a stronger effect in the improvement of 10-year stroke risk prediction, reflected by an increase in C statistic, in women compared with men. This may be explained by the presence of fewer traditional stroke risk factors and more SBI as well as subcortical and periventricular WML in women than in men. Furthermore, our findings underscore the notion that SBI and WML are not just intermediates in the association between cardiovascular risk factors and stroke, but that they may show the end organs' durability to the effects of vascular risk factors.29 The mechanisms underlying this hypothesis have not yet been elucidated and should be further investigated.
MRI is widely used in diagnostics, but it is rarely applied as a screening test. Although costly, MRI is suitable for screening because of its non-invasiveness. Accurate risk-classification is most important for individuals in the intermediate risk category as the distinction between those who will and who will not develop stroke and subsequent prevention treatments are most uncertain in this group. Moreover, people with high risk will be treated anyway, so the likely target group for application will be people with low or moderate stroke risk based on the Framingham Stroke Risk Score. There was substantial reclassification in the intermediate risk group on the basis of our new model. These results favour two-stepped screening for identification of individuals at high risk, in which people qualify for an MRI only after being classified in the intermediate risk group based on the Framingham Stroke Risk Function. This will minimise the number of people who will attend an MRI for screening and will therefore reduce costs. When we can use MRI to better distinguish people in the intermediate risk category who are at high risk of stroke from those in the intermediate category who are at low risk, earlier and more extensive preventive treatment can be focused on those who will benefit from it most. This will improve primary stroke prevention and will reduce costs considerably.
Future studies should build on this framework. Moreover, further research into other (less complex) measures that may also improve stroke prediction is clearly required, and—once a tool has been validated—further investigation should also be done on the exact target population for screening.
The strengths of this study are its population-based setting, its large number of participants and the nearly complete 10-year follow-up. Notwithstanding good-to-excellent intrareader agreement, we still may have incorrectly identified brain infarcts or misclassified infarcts as silent or symptomatic. However, SBI were identified and classified blinded for follow-up data. Misclassification may also have occurred in the identification of strokes during follow-up. General practitioners, however, have access to all available medical information of their patients and we were therefore able to identify virtually all (minor) strokes that occurred during the follow-up period (loss of potential years only 3.1%), even in participants who had not been referred to a hospital—for example, people living in nursing homes or participants who had a fatal stroke. Nevertheless, we might have missed some strokes presenting with symptoms too subtle for the participant to visit a doctor. Although this number is not likely to be high, the exact number of missed strokes is unknown. We also note that the focus of this study was on incident clinical stroke. We did not consider a TIA as an incident event and therefore also did not exclude participants with a history of TIA.
A limitation of our study is that at the time MRI scans were carried out (1995–6), manual quantification of WML volume was the state-of-the-art technique and we did not have automatic MR analyses available. We have shown that there is good-to-excellent agreement between manual and automated classification methods.30 Nonetheless, an automated classification system would be more practical in clinical settings. In addition, owing to the relatively small number of stroke subtypes (12 haemorrhagic, 59 ischaemic, 28 unspecified) we were not able to investigate prediction in strata of stroke subtypes. The strength of the associations may differ between the two subtypes, which might also have affected our results. It is important to note that the Framingham Stroke Risk Function also includes all subtypes of stroke in the model.4 Moreover, we scored the presence of brain infarcts on MRI (no/yes) without distinction in numbers. As a result, we were not able to investigate any dose–response relationship between the numbers of brain infarcts and incidence of stroke.
Another limitation is that our statistical methods did not take into account competing risks. However, as a larger burden of WML also leads to a higher death rate due to other disease,31 the associations we found may be even stronger if we had taken these competing risks into account.
To summarise, we found that assessment of small vessel disease with MRI beyond the classic stroke risk factors improved the estimation of subsequent stroke, especially in women with an intermediate stroke risk. These findings support the use of cerebral imaging as a possible tool for better identifying people at a high risk of stroke.
Funding The Rotterdam Study is supported by Erasmus Medical Center and Erasmus University Rotterdam, the Netherlands Organization for Scientific Research (NWO), the Netherlands Organization for Health Research and Development (ZonMW), the Research Institute for Diseases in the Elderly (RIDE), the Netherlands Genomics Initiative, the Ministry of Education, Culture and Science, the Ministry of Health, Welfare and Sports, the European Commission (DG XII) and the Municipality of Rotterdam.
Competing interests None.
Ethics approval Ethics approval was provided by the medical ethics committee of Erasmus MC University Medical Center.
Provenance and peer review Not commissioned; externally peer reviewed.