Page 414

Karger_ESC London_2013

22. European Stroke Conference 414 © 2013 S. Karger AG, Basel Scientific Programme Figure. A. In the real tDCS session, there was increased activation in right M1 and decreased activation in left M1 and both SMA during left hand task. B. When fMRI images between the real and sham tDCS sessions were compared, the real tDCS session showed significantly increased acti-vation in the both SMA and left M1 in motor network after applying tDCS. 242 Brain imaging Hyperacute fingerprinting: CT brain machine-learning predicts response to thrombolysis P. Bentley1, J. Ganesalingam2, A. Dias3, A. Mehta4, P. Sharma5, O. Halse6, D. Rueckert7 Imperial College London, London, UNITED KINGDOM1, Imperial College London, London, UNITED KINGDOM2, Imperial College London, London, UNITED KINGDOM3, Imperial Col-lege London, London, UNITED KINGDOM4, Imperial College London, London, UNITED KING-DOM5, Imperial College London, London, UNITED KINGDOM6, Imperial College London, Lon-don, UNITED KINGDOM7 Background: Thrombolysis is the principle treatment of acute ischemic stroke, but is associated with an adverse event rate of upto 10%, mostly due to haemorrhage that can be fatal. Being able to predict who will come to harm should improve patient selection, resulting in enhanced overall effec-tiveness. Although certain recognized hyperacute imaging features may hold prognostic value, these can be difficult to quantify, and are unlikely to be exhaustive. Here we show how pattern-recogni-tion techniques applied to hyperacute CT can help predict thrombolysis-induced hemorrhage. Methods: Data from 98 ischemic strokes treated with thrombolysis were obtained, comprising: 11 with, and 87 without, subsequent symptomatic haemorrhage. Admission CT brains were spatially normalized; 190,000 2mm3 brain voxels extracted, and intensities normalized. This provided input to a support vector machine (SVM) that classified treatment outcomes dichotomously. Performance was assessed by a cross-validation method in which patients were repeatedly split into training sets of 88, and test sets of 10, with each test set always having only one haemorrhage (550 runs). Test subjects were ranked according to likelihood of haemorrhage as estimated by the SVM, with only the 1st rank subject being a predicted haemorrhage (assuming 10% event rate). Results: Of the 11 patients who haemorrhaged, 5 were consistently and correctly predicted by the SVM as the most likely out of ten treated patients to do so (permutation test: p<1.exp-6). Important-ly, there were no features on manual inspection of any of their hyperacute scans indicative of hae-morrhage risk (e.g. low ASPECTS score). Furthermore, a logistic regression model of haemorrhage using established predictors was not significant in our sample. Conclusion: Machine learning applied to hyperacute stroke imaging can contribute to patient selec-tion for thrombolysis. Ongoing work aims to integrate imaging with clinical data and utilise CT per-fusion.


Karger_ESC London_2013
To see the actual publication please follow the link above