Abstract
Treating children with autism spectrum disorders (ASD) with behavioralinterventions, such as Pivotal Response Treatment (PRT), has shown promise inrecent studies. However, deciding which therapy is best for a given patient islargely by trial and error, and choosing an ineffective intervention results inloss of valuable treatment time. We propose predicting patient response to PRTfrom baseline task-based fMRI by the novel application of a random forest andtree bagging strategy. Our proposed learning pipeline uses random forestregression to determine candidate brain voxels that may be informative inpredicting treatment response. The candidate voxels are then tested stepwisefor inclusion in a bagged tree ensemble. After the predictive model isconstructed, bias correction is performed to further increase predictionaccuracy. Using data from 19 ASD children who underwent a 16 week trial of PRTand a leave-one-out cross-validation framework, the presented learning pipelinewas tested against several standard methods and variations of the pipeline andresulted in the highest prediction accuracy.