Fig. 8: Replication on the SRPBS dataset and validation on the HCP dataset.

a The box-plot charts of Pearson’s correlation coefficients and MAE between the actual and predicted scores using single-task learning (green) and multi-task learning (orange) over 10 runs of five-fold cross-validation on the SRPBS sample (n = 206). “*” indicates that multi-task learning significantly improves the accuracy compared to single-task learning (Student’s t-test with FDR correction via the Benjamini–Hochberg procedure, p < 0.05). b The distinct and shared importance maps derived for the SRPBS sample, where higher attention weights indicate greater contributions to the prediction of both cognitive functioning and SZ severity. c The prediction performance on a dataset of healthy subjects—HCP (n = 875). The model pretrained on the COBRE dataset achieved significantly lower MAE values—closer to zero—than the model trained from scratch. Moreover, pretraining significantly improved prediction correlations across all four cognitive domains and substantially reduced MAE for processing speed and working memory. “*” denotes these improvements (Student’s t-test with FDR correction, p < 0.05).