Study design

This retrospective observational matched‑cohort study aimed to investigate the effect of hemostasis‑efficient DOAC administration on the course and outcomes of AIS in patients hospitalized at the Department of Adult Neurology of the University Clinical Centre in Gdańsk from December 2020 to January 2024.

Our investigation was conducted and subsequently reported in compliance with the STROBE guidelines for cohort studies.¹²

We analyzed patients diagnosed with AIS, categorized into 2 cohorts. The observation (DOAC) group comprised individuals with a DOAC concentration of over 50 ng/ml on acute admission, with prescribed DOACs (dabigatran, rivaroxaban, or apixaban) before stroke, primarily for preventing cardioembolic AF‑related events; no patients were receiving DOAC therapy solely for embolic stroke of undetermined source without another established indication. The control (non‑DOAC) group included properly matched AIS patients, including patients with AF, admitted to a hospital in 2023, who did not receive DOACs prior to admission due to various reasons, such as refusal of treatment or interruption of administration before stroke long enough (>48 h) to be classified into the non‑DOAC group on admission.

Plasma DOAC concentrations were routinely measured in all AIS patients with documented prior DOAC use, irrespective of the reported time from last intake. The individuals with plasma DOAC levels equal to or lower than 50 ng/ml were excluded from the analysis, as the study focused exclusively on the cases with laboratory‑confirmed, therapeutically relevant anticoagulant activity.

The initial unmatched study population consisted of 152 participants. Their median (interquartile range [IQR]) age was 80.5 (73–85) years, and 82 were women (53.9%). Median (IQR) time from stroke onset to the first National Institutes of Health Stroke Scale (NIHSS) assessment was 10.3 (4–20) days (Table 1). In comparison with the non‑DOAC group, the patients on anticoagulant therapy had a higher prevalence of diabetes (46.7% vs 29.3%; P = 0.04), and a greater number of them had AF (93.3% vs 34.8%; P <⁠0.001).

Data collection and analysis were carried out using the MedStream Designer platform (Transition Technologies S.A., Warszawa, Poland). To streamline the process, the participants were matched based on age and sex to create the dataset for the non‑DOAC cohort.

The exclusion criteria involved receiving specific treatments for AIS (eg, mechanical thrombectomy or intravenous thrombolysis), administration of vitamin K antagonists (VKAs), therapeutic‑dose heparin, unfractionated heparin, or DOAC treatment without documented levels or in nontherapeutic doses.

During the initial preselection phase, the patients were matched in a 1.5:1 ratio using a cardinality matching algorithm,¹³ with age and sex as covariates. This ratio was chosen to establish an enriched pool of potential controls (n = 111) for the DOAC cases (n = 74). This level of oversampling was intended to serve as a buffer to address potential unmatched pairs in the subsequent phase, while ensuring that the manual data collection burden on clinicians remained manageable, as compared with 2:1 and 3:1 ratios. Subsequently, the primary matching procedure was repeated with a 1:1 ratio using the same algorithm. The primary matching process included covariates, such as AF, age, sex, prestroke modified Rankin Scale (pre‑mRS), chronic kidney disease, hypertension, diabetes, dyslipidemia, community‑acquired infection, previous IS, previous transient ischemic attack (TIA), coronary artery disease, myocardial infarction, carotid artery stenosis, and smoking.

The study received approval from the Independent Bioethics Committee for Scientific Research at the Medical University of Gdansk, Poland (KB/75‑439/2024).

Participant eligibility

We identified 74 patients with AIS on DOAC treatment and 407 AIS patients admitted in 2023, who were not taking DOACs. Individuals who were eligible to receive recanalization–reperfusion treatments were excluded. Out of 407 non‑DOAC patients, 195 were further excluded for the following reasons: using other anticoagulants (VKAs, unfractionated heparin, therapeutic‑dose heparin), no measurement of anticoagulant activity or its nontherapeutic concentration, presenting to an emergency department more than 7 days after stroke onset, and initially suspected AIS not confirmed at discharge. During the initial preselection phase, the patients were matched based on age and sex, resulting in the inclusion of 111 non‑DOAC participants. The individuals with TIA were not included in the analysis, as no anticoagulant activity was routinely tested in this group. After addressing missing data and algorithmic mismatches, the DOAC group had 60 the non‑DOAC cohort 92 participants. Supplementary material shows the study flow diagram (Figure S1) and clinical data describing the investigated cohort (Table S1).

Data

The primary end point was a functional outcome reflected in a 3‑month mRS score, obtained through telephone interviews with the patients or their caregivers (0–2 points indicated good, and 0–1 excellent outcomes). Secondary end points included 3‑month all‑cause mortality rate and the NIHSS score on day 9, which is the typical day of discharge in our institution for patients without major stroke complications. In the cases where the discharge was delayed for medical reasons, NIHSS data were collected on the actual day of discharge instead of day 9.

To enhance measurement consistency and minimize variability, we employed several strategies to reduce bias. To limit selection bias, all eligible patients meeting the inclusion criteria within the study period were initially included in the analysis. Secondly, the matching was utilized to balance key clinical and demographic variables. Finally, cross‑referencing, self‑reported data with clinical records were used to mitigate potential inaccuracies in medical history reporting. Missing data were handled through multiple imputations, although missing not at random data could still impact internal validity.

Statistical analysis

The sample size was based on patient data availability in electronic health records. The mRS scores were recorded numerically. For the analysis, binary outcomes were derived from the mRS scores, and dichotomized into several categories: 0–1 vs 2–6 points, 0–2 vs 3–6 points, and 6 vs 1–5 points.

Descriptive statistics were used to summarize patient characteristics. Quantitative variables were expressed as medians with IQRs, while categorical variables were presented as numbers and percentages. Missing values were also reported. Baseline differences between the groups were assessed using the permuted Yuen–Welch t test with 1000 replications and 10% trimming. The Fisher exact test evaluated proportion differences.

Postmatching covariate balance was assessed using standardized mean differences (SMDs), with SMD below 0.1 indicating satisfactory balance between the DOAC and non‑DOAC groups. The MatchThem package (Farhad Pishgar, Johns Hopkins University, Baltimore, Maryland, United States) was used for matching.¹⁴

The effect of DOACs on functional independence was analyzed using a generalized linear model (GLM) with binomial distribution and identity link function, which yielded unadjusted risk difference (RD) with 95% CI. This approach was applied to other binary outcomes. For NIHSS discharge scores, a GLM with the Gaussian distribution and identity link function were utilized. The outcome underwent log transformation (NIHSS + 1) to normalize the distribution and account for zero values. The unadjusted geometric mean ratio and 95% CI were computed. P values were derived using the Wald method, with significance at a P value below 0.05. No adjustments were made for multiple testing.

Missing values were handled using the MICE package (TNO Life Sciences, Leiden, the Netherlands),¹⁵ applying 20 imputations under the missing at random assumption. Model coefficients were combined using the Rubin rules. Matched cohort frequencies were omitted due to varying sample compositions across imputed datasets. Sensitivity analysis used propensity score analysis with a 0.1 caliper and greedy nearest‑neighbor matching to assess the robustness of our primary findings. Statistical analysis was conducted using R software, version 4.4.2 (R Foundation for Statistical Computing, Vienna, Austria).

Limitations

This study has several limitations. Its primary shortcoming is a lack of detailed information regarding the timing of the last DOAC intake. Secondly, we did not report events of intracranial bleeding, which is a complication potentially associated with the end points. Given that the investigation relied on data collected retrospectively, it is possible there were unmeasured and unidentified confounding variables that might have impacted the outcomes. Moreover, the retrospective data limited the likelihood of establishing causal inferences. The strict inclusion criteria and matching may have restricted the generalizability of the findings to wider populations. Finally, and importantly, the small sample size, especially with the matching procedure, might have diminished the statistical power of detecting significant effects.

Conclusions

This study demonstrated a significant reduction in stroke severity on discharge in patients with high admission plasma anticoagulant activity from DOAC administration. This improvement did not translate into a difference in functional outcomes at 90 days; however, the small sample size might have led to underestimation of the treatment effect.