Background
Recurrent stroke accounts for approximately a quarter of all strokes that occur and has important implications for the long-term outcome of patients [
1]. The 1-year incidence of recurrent ischemic stroke has been estimated to range from 8 to 14% [
2,
3]. The risk estimation of recurrent ischemic stroke can be achieved by using prediction models.
Many clinical predictors for recurrent ischemic stroke have been investigated. Strong evidence was established for a limited number of factors, which include stroke prior to the index stroke and stroke subtype [
4]. Other factors such as age, sex, hypertension, diabetes mellitus, hyperlipidemia, smoking, history of myocardial infarction, history of atrial fibrillation, and history of peripheral artery disease have been suggested as predictors of recurrent ischemic stroke in some studies, but not in others [
4].
CT is often the imaging modality of choice for diagnosing acute ischemic stroke mostly due to high availability and lack of contraindications [
5]. Previously identified imaging predictors for recurrent ischemic stroke include acute ischemia on non-contrast CT (NCCT), occlusion or stenosis on CT angiography (CTA), and poor collateral supply on CTA [
6,
7]. In addition, magnetic resonance imaging (MRI)–derived predictors such as multiple ischemic lesions are shown to have added value to clinical models [
4,
8,
9].
Clinical prediction models have been summarized in a systematic review and meta-analysis [
10]. The discriminative performance of these models was moderate [
10]. Most models are developed for predicting recurrences of ischemic stroke not longer than 90 days or 1 year after the initial stroke, whereas ischemic stroke may recur up to 5 years after the initial stroke and beyond [
11]. To our knowledge, the only model that was developed to predict 5-year ischemic stroke recurrence was developed in young stroke patients. Prediction of 5-year recurrent stroke in adult patients has not been studied before, and the added value of CT-derived predictors is unknown. Therefore, we sought to develop a model incorporating clinical risk factors for predicting 5-year recurrent ischemic stroke in patients with ischemic stroke and to determine whether adding CT-derived predictors improves prediction of recurrent ischemic stroke over 5 years of follow-up.
Discussion
In this study, addition of two CT imaging variables (previous cerebral infarcts on NCCT and ASPECTS on mean transit time maps derived from CTP) to the clinical model resulted in a significant improvement in discrimination performance for ischemic stroke recurrence over 5-year follow-up.
CT is often performed in patients with ischemic stroke as CT has some advantages over MRI in the acute stroke setting such as the short acquisition time, patient compatibility, costs, and availability. Besides diagnostic purposes, findings on admission CT can be used for prognostic purposes in patients with acute ischemic stroke. For instance, several studies investigated the value of CT findings on predicting clinical outcome after ischemic stroke [
13,
25]. The added value of predictors derived from CTP and CTA for predicting clinical outcome after 3 months was limited in the DUST dataset [
13]. A previous study showed that prediction of recurrent ischemic stroke at 90 days was significantly improved by adding MRI-derived predictors such as multiple infarcts and involvements of multiple vascular territories to a clinical model [
8]. In addition, similar MRI-derived predictors were identified as being predictive of recurrent ischemic stroke during a 2-year follow-up [
9]. However, the added value of CT-derived predictors has never been investigated for predicting recurrent ischemic stroke beyond 2 years. In this study, we showed that recurrence risk estimation can be significantly improved by adding CT-derived predictors to a model incorporating clinical predictors.
Although stroke subtype was a significant predictor of recurrent ischemic stroke in previous studies, this predictor did not have added value to the clinical prediction model in our study population [
4]. The stroke etiology was determined using the TOAST classification [
14]. Determination of the stroke etiology often requires extensive diagnostic work-up and can sometimes only be accurately determined during a follow-up period after the stroke, whereas immediate recurrence prediction after the index event is desirable [
26]. In this study, the TOAST classification was determined during the initial admission phase. As follow-up studies for determining the final stroke etiology were not routinely taken into account, the cause of the stroke remained unknown for 31% of the cases. Therefore, the results from this particular analysis need to be interpreted with caution. Future studies should elucidate whether stroke subtype has added value to long-term prediction of recurrent stroke.
The added value of previous cerebral infarcts on NCCT to a history of either stroke or TIA can be explained by the fact that brain ischemia may occur without the patient noticing, which is called a silent brain infarction. The association between silent brain infarction and future stroke has been established before, but it has never been related to recurrent ischemic stroke [
27]. Whether the patient has had a previous ischemic stroke or TIA is usually evaluated by history taking. However, it is possible that ischemic brain damage is present, although the patient has not experienced any stroke symptoms. This is a typical example of how CT imaging has added prognostic value to clinical assessments such as history taking.
Intuitively, prediction models will be more reliable if they include predictors that are already well-established risk factors for recurrent stroke. We found that poor collaterals did not seem to be an important predictor for recurrent stroke. This finding does not mean that poor collateral is not a risk factor for recurrent stroke by itself, but, instead, this factor has no added value to the prediction of recurrent stroke beyond the other predictors used in our risk prediction model. An explanation for this finding might be the lack of power due to the small number of patients with poor collaterals. This finding needs verification in a more balanced population.
The observed association between higher ASPECTS on mean transit time maps derived from CTP and increased recurrence risk is surprising, because it implies that patients with smaller areas of ischemia and/or involvement of less ASPECTS regions face a higher risk of recurrence compared with patients with greater areas of ischemia and/or involvement of more ASPECTS regions. We were not able to distinguish between the infarct size and the multiplicity of the infarct as these data were not routinely collected in the DUST. Additional studies are warranted to confirm this remarkable finding and to assess its relation to infarct size, multiplicity, and etiology. An advantage of using ASPECTS is that it can be accurately graded in the acute stroke phase and that it can be instantly used for prediction purposes. Although ASPECTS has been initially developed for NCCT assessments, it can also be applied to other CT modalities such as CTP [
15,
28]. In this study, the dichotomized measure of ASPECTS had added value to the clinical prediction model, making it a promising tool for recurrence prediction purposes. This finding however needs verification in a larger study with prospective outcome evaluation.
In this study, we showed that recurrence prediction after ischemic stroke can be improved by using imaging information in addition to clinical information. However, even after the model was improved, the performance was still moderate. Some steps need to be taken before a model that predicts recurrent ischemic stroke can be used in routine stroke care. First, studies may look for additional predictors (e.g., derived from imaging) to see if a clinically relevant improvement can be achieved. For example, current studies are also focusing on including the heart in the stroke admission scan to improve the early diagnosis of cardioembolic causes. Preferably, the found predictors such as ASPECTS on MTT maps should be validated in a separate study cohort. Second, once a model with a sufficiently high performance is developed, it needs to be validated in other cohorts. Third, ideally, the impact of the model needs to be quantified in a randomized controlled trial. In this way, the prognostic model may guide treatment decisions and therefore affect patient outcomes. This study contributes to the process of finding an optimal model for recurrence prediction.
Strengths of this study were the long-term follow-up and the selection of candidate predictors, which were based on literature. In this way, we avoided selecting predictors purely on significant
P values. In addition, predictor information was collected, prospectively leading to a minimal number of missing values. A limitation of this study was the retrospective collection of follow-up data, which could have induced underestimation of the outcome prevalence. This could have influenced our results in case certain associations are related to the loss of follow-up. For instance, recurrences, which were recorded in another hospital than the hospital of the index stroke, were missed. We do not believe that this has happened often, as most patients return to their own hospital for follow-up visits. Still, the observed prevalence is in line with previous studies, but studies with prospective follow-up are needed to verify our findings. The number of outcomes was relatively small, which was also a limitation of this study. Heart failure was not collected as a potential predictor in this study, whereas it showed to be of predictive value in previous studies [
29,
30]. However, with less than sixty recurrences, we were not allowed to add more than five predictors to our extended model. Selecting only three out of fourteen DUST centers contributed to this limitation, but acquiring follow-up data from the other DUST centers was not deemed feasible. Studying larger cohorts may allow more predictors into the final model. Instead of improving a previously developed model, we had to create our own clinical model that best fitted our data. A drawback of this method is that our clinical model needs validation in other studies, whereas a previously developed model has already been validated.
In conclusion, clinical models for predicting long-term recurrence after ischemic stroke have moderate performance and can be improved by adding CT-derived predictors.
Acknowledgments
The Dutch acute stroke study (DUST) investigators are:
Academic Medical Center, Amsterdam, The Netherlands (Majoie CB, Roos YB); Catharina Hospital, Eindhoven, The Netherlands (Duijm LE, Keizer K); Erasmus Medical Center, Rotterdam, The Netherlands (van der Lugt A, Dippel DW); Gelre Hospitals, Apeldoorn, The Netherlands (Droogh - de Greve KE, Bienfait HP); Leiden University Medical Center, Leiden, The Netherlands (van Walderveen MA, Wermer MJ); Medical Center Haaglanden, The Hague, The Netherlands (Lycklama à Nijeholt GJ, Boiten J); Onze Lieve Vrouwe Gasthuis, Amsterdam, The Netherlands (Duyndam D, Kwa VI); Radboud University Nijmegen Medical Center, Nijmegen, The Netherlands (Meijer FJ, van Dijk EJ); Rijnstate Hospital, Arnhem, The Netherlands (Kesselring FO, Hofmeijer J); St. Antonius Hospital, Nieuwegein, The Netherlands (Vos JA, Schonewille WJ); St. Elisabeth Hospital, Tilburg, The Netherlands (van Rooij WJ, de Kort PL); St. Franciscus Hospital, Rotterdam, The Netherlands (Pleiter CC, Bakker SL); VU Medical Center, Amsterdam, The Netherlands (Bot J, Visser MC); and University Medical Center Utrecht, Utrecht, The Netherlands (Velthuis BK, van der Schaaf IC, Dankbaar JW, Mali WP, van Seeters T, Horsch AD, Niesten JM, Biessels GJ, Kappelle LJ, Luitse MJ, van der Graaf Y).
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.