Early Trajectory of Stroke Severity Predicts Long-Term Functional Outcomes in Ischemic Stroke Subjects

Study Design

The ESCAPE trial was an investigator initiated multicenter, prospective, randomized, open-label controlled trial with blinded endpoint assessment (PROBE) design assessing the additional benefit of modern endovascular treatment when compared with guideline-based standard of care. Participants were assigned, in a 1:1 ratio, to receive endovascular treatment plus guideline-based care (intervention group) or guideline-based care alone (control group). The trial screened subjects fulfilling clinical eligibility criteria if they presented within 12 hours of stroke symptom onset and then included them only if they met neurovascular imaging criteria. All subjects had standard assessment for demographics, medical history, previous medications, baseline laboratory tests, and stroke severity (defined by the NIHSS score). The trial enrolled 316 subjects with 165 subjects randomized to intervention arm and 150 subjects randomized to control arm, and 1 excluded due to improper consent procedures. After randomization, NIHSS was assessed within the first 8 hours, at 24 hours, 48 hours, 5 days, 30 days, and 90 days. The primary outcome was the mRS at 90 days after randomization.^10,11

Group-based trajectory modeling (GBTM) was used to identify subject subgroups with distinct longitudinal trajectories of stroke severity, as measured by the NIHSS over a 48-hour, 5-day, and 30-day period after randomization. GBTM is a flexible methodology that models heterogeneity in longitudinal trajectory of individuals, while identifying latent (ie, unobserved) subgroups of individuals with the similar longitudinal trajectories. GBTM estimates subject-specific posterior probability of belonging to different latent subgroups. This methodology simultaneously implements several regression models through maximization of a likelihood that combines the information from all models. Specifically, the probability of belonging to each potential adherence group is modeled as a simple multinomial logistic regression with no predictors (only an intercept for each group). Within each subgroup, stroke severity is modeled as a smooth function of time using up to a fourth-order polynomial. Therefore, the model assumes that repeated observations on the same individual are independent conditional on trajectory group, meaning that the within-person correlation structure is explained completely by the estimated trajectory curve for each person’s group. The output of a group-based trajectory model includes estimated probabilities of group membership for each individual and each group and an estimated trajectory curve over time for each group.^12–16

Model selection involved the iterative process of estimating the number of trajectory groups, average probability of group membership, and the shape/order of each trajectory group using both statistical and nonstatistical considerations. Specifically, model selection begins with one quartic group, and more groups were added only if a better fit was detected using the above criteria. In addition, only the polynomial terms (quartic, cubic, or quadratic) with significant coefficients were kept before adding additional groups, but the linear terms were retained whether they were significant. The selected model is considered to be adequate if the average of the posterior probabilities of group membership for individuals assigned to each group exceed 0.7. Measures of goodness-of-fit of the GBTM are generally based on information-theoretical approaches such as the Akaike Information Criterion and Bayesian Information Criterion, with smaller values of AIC, and Bayesian Information Criterion indicating better model fit. Descriptive statistics, including means, SD, percentages, are used to summarize subject demographic, clinical, and imaging characteristics by the identified trajectory subgroups. Fisher exact test was used to assess the differences in categorical subjects’ characteristics across the 3 trajectory subgroups, whereas ANOVA was used to assess the differences in continuous subject characteristics across trajectory subgroups.

The prognostic value of the trajectory subgroups and other markers of stroke severity in predicting 90-day mRS was assessed using logistic regression analysis. Specifically, a logistic regression model¹⁷ was used assess the predictive accuracy of trajectory subgroups, baseline NIHSS, NIHSS at 24 hours, change in NIHSS between baseline and 24 hours, and infarct volume in predicting 90-day stroke outcomes with and without adjustment for other baseline characteristics including age, sex, Alberta Stroke Program Early CT Score, baseline occlusion, and intravenous alteplase. The presence of multicollinearity among the covariates was evaluated using the criterion of a variance inflation factor ≥4. For each model, odds ratios with 95% confidence intervals were used to describe the effect size of each predictor, whereas the area under the receiver operating characteristic curve (AUC) was used as a measure of predictive accuracy of the logistic regression prediction model. The Hosmer–Lemeshow test¹⁷ was used to assess the goodness-of-fit and calibration of the logistic regression models. All analyses were conducted in SAS 9.4.¹⁸