Fig. 5.

Training on controls, predicting patients. (a) Schematics for the two analyses. Above: in the leave-one-out (LOO) analysis, data for one subject was removed from the training set and used to evaluate the model prediction. The procedure was then repeated for each subject, resulting in 103 subject models and predictions. Below: we performed a one-time split of data into patient and controls groups in the transfer analysis. Using the control data, we trained a single model that was used to make predictions for all patients. Example predictions are shown (to the right of each schematic) for one subject (patient (d) from Fig. 1). The examples show that the two analyses produce qualitatively similar predictions. (b) Histograms of correlations between subject maps show greater variability in the patient group. Both in the observed task-activation maps (upper panel, red) and in the predicted task-activation maps (lower panel, blue), patients varied more than controls. The solid bars show three standard deviations from the distribution means. The results here are from the LOO analysis, because the transfer analysis applied only to patients. (c) Scatter plot showing the correlation in patients between observed and predicted task maps (dotted line shows x = y; light blue line shows the best liner fit). The predicted maps were nearly identical in LOO and transfer analyses (r = 0.98). The transfer analysis therefore predicted variability in patients virtually as well as the leave-one-out analysis did, even though the transfer-analysis was not trained on a single patient.