Background: A predictive outcome model for intracranial hemorrhage (ICH) from patient level data would be of interest in planning clinical trials. Despite decades of effort and several different types of models attempted, there is no universally accepted model. Here, we sought to apply Information Theory precepts to select the optimum set of baseline variables that can predict outcomes using data from the control arm of ATACH-II, a recently concluded ICH trial that tested the effects of control of systolic BP (SBP).
Methods: Sixty-three logistic regression models of mortality and good outcome (modified Rankin Score of 0-3 at 90 days) were generated based on all possible combination of 6 baseline variables: NIH Stroke Scale (NIHSS), Glasgow Coma Scale (GCS), age, hematoma volume (Hem Vol), intraventricular hematoma volume (IVH Vol), and SBP. Akaike Information Criterion (AIC) a measure of information content in a model was calculated. Those with the lowest AIC are considered best.
Results: ATACH-II had 405 subjects with non-lobar hemorrhages that reported all six baseline variables and long-term outcomes. The best model of good outcome was based on NIHSS, age, Hem Vol, IVH Vol (R 2 = 0.33; p <<0.0001; AIC:421.28). The best mortality model was based on NIHSS, age, IVH Vol, and SBP (R 2 = 0.18; p <<0.0001; AIC:-59.50).
Conclusion: We show that Information Theory constructs can be used to select the best set of variables to model outcomes in ICH. While we found generally better predictive ability for good outcome compared to mortality, the best model predicted only 57% of the outcome variance. While additional types of data fitting might yield additional predictive value, the relatively modest results seen here suggest that there remain additional factors related to outcome not typically measured in ICH trials. Additional data points from other trails are being pooled to see if predictive ability improves.