What components were integrated in the multi-modal multi-task (M3T) model that demonstrated the best predictive discrimination?