on behalf of the German Stroke Study CollaborationBackground and Purpose-To date, no validated, comprehensive, and practicable model exists to predict functional recovery within the first hours of cerebral ischemic symptoms. The purpose of this study was to externally validate 2 prognostic models predicting functional outcome and survival at 100 days within the first 6 hours after onset of acute cerebral ischemia. Methods-On admission to a participating hospital, patients were registered prospectively and included according to defined criteria. Follow-up was performed 100 days after the event. With the use of prospectively collected data, 2 prognostic models were developed and internally calibrated in 1079 patients and externally validated in 1307 patients. By means of age and National Institutes of Health Stroke Scale (NIHSS) score as independent variables, model I predicts incomplete functional recovery (Barthel Index Ͻ95) versus complete functional recovery, and model II predicts mortality versus survival. Results-In the validation data set, model I correctly predicted 62.9% of the patients who were incompletely restituted or had died and 83.2% of the completely restituted patients, and model II correctly predicted 57.9% of the patients who had died and 91.5% of the surviving patients. Both models performed better than the treating physicians' predictions made within 6 hours after admission. Conclusions-The resulting prognostic models are useful to correctly stratify treatment groups in clinical trials and should guide inclusion criteria in clinical trials, which in turn increases the power to detect clinically relevant differences.

Our study provides a validated prognostic model for prediction of complete recovery following ICH which could be very useful for the design of clinical studies.

Background and Purpose-Increased sympathetic drive after stroke is involved in the pathophysiology of several complications including poststroke immunudepression. β-Blocker (BB) therapy has been suggested to have neuroprotective properties and to decrease infectious complications after stroke. We aimed to examine the effects of random pre-and onstroke BB exposure on mortality, functional outcome, and occurrence of pneumonia after ischemic stroke. Methods-Data including standard demographic and clinical variables as well as prestroke and on-stroke antihypertensive medication, incidence of pneumonia, functional outcome defined using modified Rankin Scale and mortality at 3 months were extracted from the Virtual International Stroke Trials Archive. For statistical analysis multivariable Poisson regression was used. Results-In total, 5212 patients were analyzed. A total of 1155 (22.2%) patients were treated with BB before stroke onset and 244 (4.7%) patients were newly started with BB in the acute phase of stroke. Mortality was 17.5%, favorable outcome (defined as modified Rankin Scale, 0-2) occurred in 58.2% and pneumonia in 8.2% of patients. Prestroke BB showed no association with mortality. On-stroke BB was associated with reduced mortality (adjusted risk ratio, 0.63; 95% confidence interval, 0.42-0.96). Neither prestroke BB nor on-stroke BB showed an association with functional outcome. Both prestroke and on-stroke BB were associated with reduced frequency of pneumonia (adjusted risk ratio, 0.77; 95% confidence interval, 0.6-0.98 and risk ratio, 0.49; 95% confidence interval, 0.25-0.95). Conclusions-In this large nonrandomized comparison, on-stroke BB was associated with reduced mortality. Prestroke and on-stroke BB were inversely associated with incidence of nosocomial pneumonia. Randomized trials investigating the potential of β-blockade in acute stroke may be warranted.

In the last 15 years several machine learning approaches have been developed for classification and regression. In an intuitive manner we introduce the main ideas of classification and regression trees, support vector machines, bagging, boosting and random forests. We discuss differences in the use of machine learning in the biomedical community and the computer sciences. We propose methods for comparing machines on a sound statistical basis. Data from the German Stroke Study Collaboration is used for illustration. We compare the results from learning machines to those obtained by a published logistic regression and discuss similarities and differences.