Fig. 6: The unique importance maps contribute to illness severity for the COBRE and IMH samples.

a The unique importance maps for SZ severity, capturing the difference between the overlapping important regions across the four PANSS subscales and total scores from top-K pooling layers and the shared importance map from the shared attention model. A higher importance weight indicates a greater contribution to prediction. b, c The probability that regions contribute to illness severity at the regional level and modular levels, as confirmed by meta-analysis.