Effect of sample size on multi-parametric prediction of tissue outcome in acute ischemic stroke using a random forest classifier

The tissue outcome prediction in acute ischemic stroke patients is highly relevant for clinical and research purposes. It has been shown that the combined analysis of diffusion and perfusion MRI datasets using high-level machine learning techniques leads to an improved prediction of final infarction compared to single perfusion parameter thresholding. However, most high-level classifiers require a previous training and, until now, it is ambiguous how many subjects are required for this, which is the focus of this work. 23 MRI datasets of acute stroke patients with known tissue outcome were used in this work. Relative values of diffusion and perfusion parameters as well as the binary tissue outcome were extracted on a voxel-by-voxel level for all patients and used for training of a random forest classifier. The number of patients used for training set definition was iteratively and randomly reduced from using all 22 other patients to only one other patient. Thus, 22 tissue outcome predictions were generated for each patient using the trained random forest classifiers and compared to the known tissue outcome using the Dice coefficient. Overall, a logarithmic relation between the number of patients used for training set definition and tissue outcome prediction accuracy was found. Quantitatively, a mean Dice coefficient of 0.45 was found for the prediction using the training set consisting of the voxel information from only one other patient, which increases to 0.53 if using all other patients (n=22). Based on extrapolation, 50-100 patients appear to be a reasonable tradeoff between tissue outcome prediction accuracy and effort required for data acquisition and preparation.