Early recurrence after acute ischaemic stroke in patients with non-valvular atrial fibrillation: CHA2DS2-VASc score as determinant of oral anticoagulation timing

J. Simon
M. Niewada

Background
Non-valvular atrial fibrillation (NVAF)-related acute ischaemic stroke (AIS) is reported to account for 13-26% of the global AIS burden[1]. The AIS risk of NVAF patients is almost fivefold greater than that of their normal sinus rhythm counterparts[2]. Among the cardiovascular (CV) contributors to stroke investigated in the seminal Framingham Heart Study - hypertension, ischaemic heart disease, heart failure, atrial fibrillation - AF was unique in that its effect on risk of AIS did not abate with advancing age[3]. In fact, AF was the sole CV condition to exert an independent effect on stroke incidence among patients aged 80-89 years. As such, lifelong oral anticoagulation introduced on the basis of continuing NVAF-associated thromboembolic risk has proven to be highly effective for secondary stroke prevention. In particular, oral anticoagulants (OACs) minimise the 0.5-1.3% risk of recurrence per day within the first 14 days after the index stroke, whence the risk of both early recurrence and haemorrhagic transformation (HT) is highest[4]. That AIS increases the risk of thromboembolism and haemorrhage in tandem during this 2-week period can be cogently explained as follows: matrix metalloproteinases degrade the ultrastructure of the cerebral microvasculature very early during focal cerebral ischaemia[5] such that the blood-brain barrier is disrupted and haemorrhagic transformation occurs. In turn, haemorrhage can precipitate thromboembolic lesions by semi-solidifying to parenchymal hematomas. The clinical challenge in OAC (re)introduction timing lies in preventing thromboembolic stroke recurrence whilst avoiding intracranial haemorrhage (ICH). Complicating the decision to administer or withhold OACs are the judgments attached to predicted outcomes: (a) the rate of recurrence is severalfold higher than the rate of ICH such that the probability of recurrence in the absence of OACs is greater than the probability of ICH in the presence of OACs[6], however, (b) ICH is generally associated with higher mortality and morbidity than recurrence[7].
The concern that early OAC (re)introduction may cause ICH or exacerbate HT governs much of present-day practice, prompting neurologists to delay OAC pharmacotherapy, particularly in patients with severe stroke and large areas of infarction. However, the suspicion that early OACs promote HT and, further, the assumption that HT is clinically significant and thus merits to dictate OAC timing are based on expert opinion rather than empirical evidence. Highly evocative of the absence of consensus was an online survey conducted among UK stroke physicians which reported that 95% of respondents were uncertain about optimal timing for (re)introduction of direct OACs (DOACs). Indeed, the guidelines to which clinicians refer are highly inconsistent: where the European Heart Rhythm Association, European Society of Cardiology (EHRA-ESC) introduced a stroke severity-tiered timeline known as the ‘1-3-6-12 days rule’ in 2013, the American Heart Association/American Stroke Association (AHA/ASA) recommended (re)introducing OACs 4-14 days after the onset of neurological symptoms in 2018. In contrast, the UK recommended in 2016 that OACs be deferred at least 14 days from AIS onset in patients with symptoms of disabling stroke with the addendum that clinicians could (re)introduce OACs earlier for patients with a milder symptomatology, at their discretion. Still more, a German guideline (2015) stated that the efficacy of DOACs has not been established within < 14 days of stroke but does not draw a directive from this statement. Importantly, while Japan’s guidelines (2011) predated the use of DOACs, none of the aforelisted guidelines distinguish between the timing of VKAs versus DOACs. This is problematic given the very substantial differences in their pharmacodynamics, including time to achieve INR ≥ 2.0 - 2-4 days for a VKA compared to < 2 hours for the DOAC dabigatran - and duration of action. Parenthetically, dabigatran treatment was reported to begin sooner than any other DOAC, particularly for patients with milder disease states who were capable of taking dabigatran etexilate mesylate capsules. This finding concurs with larger studies that report OAC (re)introduction as early as during the acute phase of AIS in mild AIS patients, suggesting that moderate-tosevere AIS patients may be particularly disadvantaged by current practice. Of note, early (re)introduction of OACs, especially when the patient is still in the hospital, may be advantageous beyond the reduction of recurrence by improving compliance, expediting discharge, and reducing healthcare costs.
Overall, the decision of when to (re)introduce OACs post-NVAF-related AIS remains the object of individual, often arbitrary clinical judgment, founded only on expert opinion and observational data that constitute level C evidence. The lack of high-quality evidence is a consequence of the routine exclusion of AIS patients from phase III randomised clinical trials in the context of stroke prevention in atrial fibrillation (SPAF). In order to minimise the risk of HT, patients were excluded for at least 7 days after AIS in the ARISTOTLE trial (apixaban), at least 14 days after minor AIS and 90 days after major AIS in the ROCKETAF trial (rivaroxaban), and at least 14 days after minor AIS and 180 days after major AIS in the RE-LY trial (dabigatran). Three non-randomised, prospective observational studies, on the other hand, included patients with recent NVAF-related AIS when exploring DOAC use. The studies reported annualised risks of recurrence ranging from 7.7 to 8.5% per year and two of the three studies found that this risk increased when DOACs were (re)introduced late. Encouragingly, the observed rates of recurrence fell within the confidence intervals of estimated recurrence risk corresponding to their cohorts’ median baseline CHA2DS2-VASc scores. Moreover, two of the three studies reported low annualised risks of symptomatic ICH (sICH, 0.9 and 1.3%) with early DOAC (re)introduction, while one reported a considerably higher risk of 6.4% although the association between early DOAC use and sICH was uncertain because the haemorrhagic event occurred after 30 days. Generally, these preliminary findings have been met with optimism and four randomised controlled trials comparing early to later DOAC (re)introduction in NVAF-related AIS are underway, collectively recruiting up to 9000 patients.
The aforementioned CHA2DS2-VASc score is a well-validated clinical prediction rule (CPR) that estimates the annual risk of ischaemic stroke in patients with NVAF on the basis of the following comorbid features: congestive heart failure (CHF, C), hypertension (H), diabetes mellitus (DM, D), peripheral vascular disease (V), age 65-74 years (A), and sex category (female, Sc), each assigned 1 point, and age ≥75 years (A2) and antecedent episodes of stroke or transient ischaemic attack (S2), each assigned 2 points. CHA2DS2- VASc is the CPR most commonly used in clinics to guide decisions about anticoagulation therapy in AF patients[3]. Beyond risk of index or recurrent stroke, CHA2DS2-VASc has been associated with several aspects of stroke outcome including early neurologic deterioration and mortality in patients with NVAFrelated AIS[4]. We herein focus on the value of CHA2DS2-VASc as tool to stratify patients per risk of early recurrence in the absence of OACs. Indeed, we envisage that identification of specific subgroups that are at low risk of recurrence could facilitate the clinician’s decision to cautiously delay OACs. Similarly, subgroups that are at high risk of recurrence could justify opting for very early OAC (re)introduction. The pragmatic appeal of a tool that could streamline OAC clinical protocols on the basis of a score that is reductionist, easy to remember, and simple to calculate cannot be overstated. On behalf of the VISTA collaborators, Tu et al. reported that pre-stroke CHADS2 and CHA2DS2- VASc scores were both independently associated with functional outcomes, selected acute cardiovascular events, and mortality within the first 3 months of AIS. The pre-stroke CHA2DS2-VASc score in particular proved useful in identifying patients at low risk for poor outcomes and cardiac complications within the first 3 months following AIS in both patients with and without NVAF. Nevertheless, the study stated that both CPRs had modest precision in estimating the probability of functional outcomes and 3-month mortality following AIS in both patients with and without AF. The authors further commented that neither score is sufficiently accurate for predicting functional outcomes and 3-month mortality in a given AIS patient. While Tu et al.’s study boasts large sample size, comparative groups of patients with and without NVAF, and a population that was not limited to one diagnosis (e.g. ACS, CHF), it did not acknowledge that two numerically identical CHA2DS2-VASc scores are not necessarily made equal. For example, a CHA2DS2- VASc score of 2 may be comprised of history of stroke/TIA (assigned 2 points) or, alternatively, of two individual risk factors assigned 1 point each (e.g. age 70 and hypertension). In the former scenario, a score of 2 (resulting from a positive stroke/TIA history alone) yields an estimated stroke risk of 2.5-4.5% per year, which is probably too low[18]. This example is illustrative of a research avenue that could greatly improve the prediction precision of CHA2DS2-VASc and, with that, enhance the application of the score when timing OAC (re)introduction.
Only one study has, to the best of our knowledge, analysed stratum-specific relative risk of death associated with unique combinations of C, H, A, D, and S in the older CHADS2 score[11]. By examining the predictive value of each component of the CHA2DS2-VASc score and their unique combinations, our analysis aims to decisively inform ways in which to better calibrate CHA2DS2-VASc to the reality of recurrence risk, much like the aforedescribed example led to the suggestion that S3 (where history of stroke/TIA is assigned 3 rather than 2 points) is a better fit to available data than S2[18]. Importantly, while several stroke registries possess large volumes of patient data, VISTA uniquely befits our study design in that the very high quality of data minimises the risk of samples being incomplete for one or more of the C, H, A, D, S, V, and Sc features (see statistical analysis in methodology).

(1) CHA2DS2-VASc scores can be used to stratify patients per risk of early AIS recurrence in the absence of OACs such that thresholds for low, moderate, and high risk can be defined for the following timepoints: (i) day 7, (ii) day 14, (iii) day 28, and (iv) day 90. Risk stratum is clinically significant and may subsequently guide clinicians in the decision to administer or delay OACs for a given NVAF-related AIS patient;
(2) CHA2DS2-VASc features do not carry uniform importances in determining early AIS recurrence risk. Rather, a hierarchy of importance exists among the features (C, H, A, D, S, V, and Sc) such that its quantification will yield a CHA2DS2-VASc scoring system that is better calibrated to the reality of recurrence risk. Moreover, the hierarchy may be different for early (<14 days) versus 3-month recurrence;
(3) Unique combinations of CHA2DS2-VASc features do not carry uniform importances in determining early AIS recurrence risk. Rather, the importance of a given unique combination may exceed the additive importances of its individual features such that distinctly synergistic relationships may be identifiable;
(4) In quantifying the hierarchy that exists among the CHA2DS2-VASc features and feature combinations’ importances in determining early AIS recurrence risk, a revised and improved scoring system can be obtained for CHA2DS2-VASc. This may surpass the modest predictive power of the existing scoring system for short-term stroke outcomes in AIS patients with (and without) NVAF by, among other revisions, distinguishing between scores previously considered numerically identical.

Data source
• Extraction of anonymous patient data from AIS trials in VISTA that document:
• Age,
• Gender,
• Pre-stroke modified Rankin Scale (mRS) score,
• Treatment with IV tissue plasminogen activator (tPA),
• History of CHF, hypertension, DM, stroke/TIA, vascular disease (prior MI, peripheral artery disease, or aortic plaque), and AF,
• AIS subtype per TOAST classification,
• Baseline National Institutes of Health Stroke Scale (NIHSS) score,
• 3-month mRS score,
• Health outcome records screened for stroke recurrence,
• Mortality.

Study population
We will include data from patients who received placebo investigative treatment in the source trials. Data from patients who received active investigative treatments from source trials that reported no significant effect on stroke outcomes are also included.

Study design
The CHA2DS2-VASc scores for each patient will be calculated from collated baseline data. The outcome variables include all-cause mortality, functional outcomes (quantitated per the mRS), and stroke recurrence (encompassing any episode of stroke or TIA) within the first 3 months.
Patients will be classified into two groups according to AF status. Further subdivision on the basis of stroke subtype (e.g. lacunar infarct, large artery atherosclerosis) of index stroke and recurrence stroke will be performed for a sub-analysis.
We thus analyse short-term stroke recurrence, neurological deterioration, and survival in AIS patients with and without AF with different CHA2DS2-VASc scores. Among patients with numerically equivalent CHA2DS2-VASc scores, we analyse the aforelisted short-term outcomes in patients admitted for AIS with different combinations of C, H, A, D, S, V, and Sc. The unique combined effects of different features of the CHA2DS2-VASc score on stroke recurrence, neurological deterioration, and mortality are examined to tentatively establish a point system that better captures the reality of these 3-month outcomes in a given ischaemic stroke patient. Separate models are constructed to examine the effects of C, H, A, D, S, V, and Sc on stroke recurrence, neurological deterioration, and mortality. Each feature thereby receives a coefficient that is proportionate to the feature’s effect on the outcomes of interest. These models also examine the importance of unique feature combinations. This may reveal synergistic relationships that prompt us to introduce, in addition to the new coefficients, conditional rules into the scoring system such that the coefficient assigned for ‘feature A and feature B’ exceeds the additive coefficients of ‘feature A’ and ‘feature B’.
Specifically, we use logistic regression analysis to obtain odds-ratio (OR) which will report the effect each feature has on the risk of a positive outcome (e.g. stroke recurrence), followed by using permutationbased feature importance to calculate the coefficients that CHA2DS2-VASc features merit for optimal predictive performance. To obtain the importance of each feature, a model (e.g. logistic regression model) is first trained and evaluated on a full dataset. Then, the dataset is modified by randomly shuffling the features of interest among the samples, and a new model is trained and evaluated on the modified dataset. Since the shuffling will destroy any correlation the features of interest have with the outcome, the features can no longer be used by the model. If the shuffled features were important in deciding the outcome, the performance of the model will drop significantly, whereas an unimportant feature will have a negligible effect on the performance. This way, the importance of every variable can be found. The method enables a direct comparison of feature importance between different classification models, as it is independent of the model in use. Importantly, the feature importance only indicates how important a feature is in determining the outcome; not what effect each feature will have on the outcome (i.e. whether the presence of a feature will increase or decrease the risk of stroke recurrence). To infer the effect each feature has on the outcome, the OR from the logistic regression model is used. A random forest model is also constructed, likewise analysed through permutation-based feature importance, to improve robustness of findings. An additional reason for employing a random forest model is its ability to deal with correlated features: while a logistic regression model tends to prioritise one of several correlated features and assign small coefficients - and thereby low importance - to the remainder, random forest models treat correlated features equally. Thus, if correlation between features is shown to exist and the logistic regression model assigns low importance to some of these features while the random forest predicts high importance, we rely on the random forest result.

Statistical analysis
1. Preprocessing
1.1. Data will be scanned for completeness. VISTA’s very high quality of data uniquely befits our study as the need for imputation is removed. Anticipating a negligible fraction of data to be missing (<5% of patients lacking data for one or more of the C, H, A, D, S, V, and Sc features of interest), the incomplete data points will simply be discarded. We thus do not expect to opt for the HotDeck imputation algorithm which is of use in the event of intermediate fractions of incomplete data points (e.g. 5<x<20%).
1.2. Data will be checked for co-linearity by calculating variation inflation factor (VIF). If the VIF value is high (high co-linearity), we will use principal component analysis (PCA) to extract the most important features. Alternatively, removal of one of the highly co-linear features may be preferable. If VIF is low, all features will be used as is.
1.3. Study population will be divided on the basis of AF status and further on the basis of stroke subtype at index and recurrence strokes. Caution will be taken to ensure sufficient data is available for statistical significance in every subgroup.

2. Model construction and evaluation:
2.1. We will construct a logistic regression model and a random forest model. Both models are
trained on a training set (80% of original data) using k-fold cross validation and evaluated on a test set (20%). The random forest model will be trained using a bagging technique, and the number of decision trees will be decided based on a line search to obtain the best performance. Each tree is trained using the CART algorithm. The Python module scikit-learn is used for statistical analysis.
2.2. Sensitivity, specificity, AUC, and, in case of imbalanced data, precision-recall (PR) AUC and F1-score will be used as performance metrics.
2.3.1. If the aforelisted scores indicate good models, we will use permutation-based feature importance on both models to inform a new weighting (coefficient) system for the CHA2DS2-VASc features. This method is also used to find the importance of combined features.
2.3.2. If the aforelisted scores indicate poor models (2.2.), we will reconsider the preprocessing and postprocessing steps and use regularisation techniques to improve the models.

3. Interpretation:
The individual feature and feature combination importances obtained with logistic regression and random forest analyses are directly compared to see whether the same conclusions can be drawn from both models. If the models yield similar feature and feature combination importances, a weighted mean of the importance scores will be calculated. The weighting will be based on model performance; specifically, the importance score is weighted by |AUC-0.5|/0.5, where |x| is the absolute value of x. This will result in a weighting between 0 and 1, where 0 is obtained when the AUC is 0.5 (i.e. the model predicts randomly), and 1 when the AUC is 1 or 0 (i.e. the model always predicts correctly, or the model always predicts incorrectly, which, if flipped, yields a perfect predictor). The new CHA2DS2-VASc scores are then chosen such that the proportions between them are similar to those of the feature importances. If significant synergistic relationships are revealed, we may introduce a conditional rule. For example “if only feature A is present, assign 1 point, if only feature B is present, assign 1 point, if both feature A and feature B are present, give a score of 3”. On the other hand, if a large discrepancy exists between the feature and feature combination importances generated by the models, the strengths and weaknesses of these models will be tabulated and compared to identify potential causes for such a difference. Ultimately, using the feature and feature combination weightings obtained by the models, a revised and improved scoring system is proposed for CHA2DS2-VASc. As a last step, we will test the predictive performance of our scoring system by looking at the probability of a positive outcome (e.g. stroke recurrence) for a specific numerical score and by ensuring that an increasing score is associated with an increased risk of positive outcome. We will also look at variance among the study population for a given numerical score. If this variance is smaller than that associated with the original CHA2DS2-VASc score, we conclude that our scoring system has greater predictive power. We thereby inform future studies that intend to equip neurologists and other clinicians with a CPR that is applicable to AF and non-AF patients alike, reductionist, simple to calculate, and, above all, highly predictive of short-term stroke recurrence, neurologic deterioration, and mortality after AIS.

This is an academic project, with no commercial sponsorship.

[1] Gage, B., Waterman, A., Shannon, W., Boechler, M., Rich, M., & Radford, M. (2001). Validation of clinical classification schemes for predicting stroke: results from the national registry of atrial fibrillation. ACC Current Journal Review, 10(6), 20– 21. https://doi.org/10.1016/s1062-1458(01)00458-5
[2] Wallace, E., Dillon, C., Dimitrov, B., Fahey, T., & Keogh, C. (2011). Validation of the CHADS2 clinical prediction rule to predict ischaemic stroke. Thrombosis and Haemostasis, 106(09), 528–538. https://doi.org/10.1160/th11-02-0061
[3] Saliba, W., Gronich, N., Barnett-Griness, O., & Rennert, G. (2016). The role of CHADS2 and CHA2DS2-VASc scores in the prediction of stroke in individuals without atrial fibrillation: a population-based study. Journal of Thrombosis and Haemostasis, 14(6), 1155–1162. https://doi.org/10.1111/jth.13324
[4] Ntaios, G., Lip, G. Y. H., Makaritsis, K., Papavasileiou, V., Vemmou, A., Koroboki, E., Savvari, P., Manios, E., Milionis, H., & Vemmos, K. (2013). CHADS2, CHA2DS2-VASc, and long-term stroke outcome in patients without atrial fibrillation. Neurology, 80(11), 1009–1017. https://doi.org/10.1212/wnl.0b013e318287281b
[5] Welles, C. C., Whooley, M. A., Na, B., Ganz, P., Schiller, N. B., & Turakhia, M. P. (2011). The CHADS2 score predicts ischemic stroke in the absence of atrial fibrillation among subjects with coronary heart disease: Data from the Heart and Soul Study. American Heart Journal, 162(3), 555–561. https://doi.org/10.1016/j.ahj.2011.05.023
[6] Wolf, P. A., D’Agostino, R. B., Belanger, A. J., & Kannel, W. B. (1991). Probability of stroke: a risk profile from the Framingham Study. Stroke, 22(3), 312–318. https://doi.org/10.1161/01.str.22.3.312
[7] Chien, K. L., Su, T. C., Hsu, H. C., Chang, W. T., Chen, P. C., Sung, F. C., Chen, M. F., & Lee, Y. T. (2010). Constructing the Prediction Model for the Risk of Stroke in a Chinese Population. Stroke, 41(9), 1858–1864. https://doi.org/10.1161/strokeaha.110.586222
[8] Lip, G. Y., Lin, H. J., Chien, K. L., Hsu, H. C., Su, T. C., Chen, M. F., & Lee, Y. T. (2013). Comparative assessment of published atrial fibrillation stroke risk stratification schemes for predicting stroke, in a non-atrial fibrillation population: The Chin-Shan Community Cohort Study. International Journal of Cardiology, 168(1), 414–419. https://doi.org/10.1016/j.ijcard.2012.09.148
[9] Chan, Y. H., Yiu, K. H., Lau, K. K., Yiu, Y. F., Li, S. W., Lam, T. H., Lau, C. P., Siu, C. W., & Tse, H. F. (2014). The CHADS2 and CHA2DS2-VASc scores predict adverse vascular function, ischemic stroke and cardiovascular death in high risk patients without atrial fibrillation: Role of incorporating PR prolongation. Atherosclerosis, 237(2), 504–513. https:// doi.org/10.1016/j.atherosclerosis.2014.08.026
[10] Ji, C., Wu, S., Shi, J., Huang, Z., Chen, S., Wang, G., & Cui, W. (2020). Baseline CHADS2 Score and Risk of Cardiovascular Events in the Population Without Atrial Fibrillation. The American Journal of Cardiology, 129, 30–35. https://doi.org/10.1016/j.amjcard.2020.05.035
[11] Henriksson, K. M., Farahmand, B., Johansson, S., ÅSberg, S., Terént, A., & Edvardsson, N. (2010). Survival after stroke — The impact of CHADS2 score and atrial fibrillation. International Journal of Cardiology, 141(1), 18–23. https://doi.org/10.1016/j.ijcard.2008.11.122
[12] Tu, H. T., Campbell, B. C., Meretoja, A., Churilov, L., Lees, K. R., Donnan, G. A., & Davis, S. M. (2013). Pre-Stroke CHADS2 and CHA2DS2-VASc Scores Are Useful in Stratifying Three-Month Outcomes in Patients with and without Atrial Fibrillation. Cerebrovascular Diseases, 36(4), 273–280. https://doi.org/10.1159/000353670
[13] Poçi, D., Hartford, M., Karlsson, T., Herlitz, J., Edvardsson, N., & Caidahl, K. (2012). Role of the CHADS 2 Score in Acute Coronary Syndromes. Chest, 141(6), 1431–1440. https://doi.org/10.1378/chest.11-0435
[14] Mitchell, L. B., Southern, D. A., Galbraith, D., Ghali, W. A., Knudtson, M., & Wilton, S. B. (2014). Prediction of stroke or TIA in patients without atrial fibrillation using CHADS2and CHA2DS2-VASc scores. Heart, 100(19), 1524–1530. https:// doi.org/10.1136/heartjnl-2013-305303
[15] Melgaard, L., Gorst-Rasmussen, A., Lane, D. A., Rasmussen, L. H., Larsen, T. B., & Lip, G. Y. H. (2015). Assessment of the CHA2DS2-VASc Score in Predicting Ischemic Stroke, Thromboembolism, and Death in Patients With Heart Failure With and Without Atrial Fibrillation. JAMA, 314(10), 1030. https://doi.org/10.1001/jama.2015.10725
[16] Wolsk, E., Lamberts, M., Hansen, M. L., Blanche, P., Køber, L., Torp-Pedersen, C., Lip, G. Y. H., & Gislason, G. (2015). Thromboembolic risk stratification of patients hospitalized with heart failure in sinus rhythm: a nationwide cohort study. European Journal of Heart Failure, 17(8), 828–836. https://doi.org/10.1002/ejhf.309
[17] Goto, S., Bhatt, D. L., Röther, J., Alberts, M., Hill, M. D., Ikeda, Y., Uchiyama, S., D’Agostino, R., Ohman, E. M., Liau, C. S., Hirsch, A. T., Mas, J. L., Wilson, P. W., Corbalán, R., Aichner, F., & Steg, P. G. (2008). Prevalence, clinical profile, and cardiovascular outcomes of atrial fibrillation patients with atherothrombosis. American Heart Journal, 156(5), 855–863.e2. https://doi.org/10.1016/j.ahj.2008.06.029
[18] Rothwell P. M. (2008). Prognostic models. Practical neurology, 8(4), 242–253. https://doi.org/10.1136/jnnp.2008.153619
Additional ref. Wang, H., Tian, Y., Guo, Y., Wang, Y., & Lip, G. Y. H. (2016). Multiple risk factors and ischaemic stroke in the elderly Asian population with and without atrial fibrillation. Thrombosis and Haemostasis, 115(01), 184–192. https:// doi.org/10.1160/th15-07-0577