Abstract
Abstract only
Objective:
There are several challenges in implementing models for predicting stroke or stroke related outcomes. Most of these models have average concordance, and several of the important variables cannot be modified. In this study, we updated and performed a meta-analysis of commonly utilized models to predict stroke related outcomes. Our primary aim was to evaluate the discriminative ability of the concordance statistic by adding additional studies.
Methods:
Studies reporting c-index and SE (or 95% CI) for predicting stroke or related outcomes were included in our analysis. In addition to the c-index, total participants, year of publication, type of analytical method (survival, logistic regression, neural network, etc.) and type of outcome (predicting stroke or composite outcome) were utilized. Combined effect sizes with the random model, test for heterogeneity, and publication bias were considered. Egger’s test was used to assess funnel asymmetry.
Results:
Twenty-seven models were included (patients= 1762461; c-index=14, Harrell’s c-index= 13; only stroke =21, composite=6) in the analysis. Combined mean c-index was 0.76 (95% CI: 0.71, 0.81; 95% predictive interval: 0.59, 0.93). Combined mean Harrell’s c-index was 0.65 (95% CI: 0.61, 0.69; 95% predictive interval: 0.56, 0.74). Test of heterogeneity showed high variation between studies reporting c-index and Harrell’s c-index (I2=97.49% and 80.0% respectively). Egger’s test intercept was -2.1 (95% CI: -7.2, 3.0, P > .40) for c-statistic and 1.2 (95% CI: -1.2, 3.5, P > .32) for Harrell’s c-index studies.
Conclusion:
Current studies have not improved the prediction interval significantly as compared to our previous meta-analysis for predictive or explanatory model available for stroke risk. However, recent studies were found to be more inclusive of non-traditional biomarkers (e.g., genetic, or polygenic scores) and utilized various machine learning methods that were not used before.