摘要：An exact prediction of system-specific recovery after stroke is necessary to providerehabilitation therapy based on the individual needs. (Individualised rehabilitation in recovery fromaphasia requires reliable prediction.) Prediction from lesion sites/anatomy is difficult as functions aslanguage emerge from an interaction in networks and, reorganisation is a highly dynamic process,which influences activation patterns as well as white matter integrity throughout the brain. Recoveryand treatment success depends on individual poststroke anatomy. Functional scales have verydifferent metrices and the relation to brain function is essentially unknown. Multiparametric statisticsare needed to extract information from theis high dimensional data. In a previous study (Saur et al2010) we used classification with applied support vector machines (SVM) to predict individuallanguage recovery in 21 aphasic stroke patients. fMRI data activation pattern early after strokeenabled a correct prediction in 76% of patients. When age and functional scales (language recoveryscore derived from a principal component analysis of subtests from the AAT or AABT) wereincluded/added correct assignments improved up to 86%. Thus, the application of multivariatemachine learning techniques to early fMRI and clinical and demographical data has a high potentialto improve individual prediction. Using SVM it is also possible to identify “ recovery beneficialactivity” in individual subjects, thus providing a tool for a targeted intervention for stimulationtechniques.
