Fig. 11.

The figure shows the differences in overall accuracy as well as individual class accuracies in the posttraumatic stress disorder (PTSD) dataset which can be attributed to age range differences in training/validation and hold-out test data. These consensus accuracies were obtained by combining the predictions of all the 19 classifiers in a probabilistic way to vote on the hold-out test dataset. As expected, the split wherein the training and hold-out test data were matched for age had the best performance. The accuracy in the split where the age range was different for training/validation and hold-out test data was terrible with all observations being classified as PTSD in the binary classification scenario. This shows that smaller datasets with homogeneity overestimate the actual predictive capability of the classifiers and do not generalize well to the overall population. PCS, post-concussion syndrome