Figures and data
Model trained on functional connectivity maps acquired at rest predicts (a) episodic memory (EM) of the test dataset. The models trained on movie-watching (b) but not n-back (c) datasets predicted the EM scores of the test dataset. Table 1 summarizes the p values and correlations for each model. Test datasets were obtained from the same cohort for rest, movie-watching, and n-back. The winning model trained on EM at rest was evaluated on the external COBRA cohort and yielded a significant prediction of EM scores (d). Bootstrap distributions of correlations between predicted and actual EM scores indicated no significant difference in the predictive power of EM between models trained on resting state and movie-watching data (e). Additionally, the bootstrap distribution revealed that models trained on resting state and movie-watching data yielded higher correlations than those trained on n-back data (e). Visualization of features contributing to successful prediction of EM at rest (f). A grad-CAM-derived saliency map displays the features that contributed to the model’s predictions. The hot spots overlaid on the FC map demonstrate noticeable cross-correlation contributions in “default mode” (DMN) regions. Another important feature visualized by Grad-CAM includes off-diagonal hot spots reflecting inter-connections of the DMN – “subcortical” node.
Correlation results for episodic memory (EM) score predictions
Correlation results for working memory (WM) score predictions.
Model trained on FC maps acquired at rest did not significantly predict (a) the working memory (WM) in the test dataset. The model trained on the movie-watching dataset yielded the best-performing model in predicting WM (b) while the model trained on the n-back dataset (c) was the second-best model. Table 2 summarizes the p values and correlation power for each model. (d) Results of cross-dataset validation, where the best-performing model in the DyNAMiC dataset (i.e., movie-watching) was applied to predict WM to the COBRA dataset. However, since COBRA does not include a movie-watching paradigm, we applied the model to the n-back task in COBRA(Table 2). Bootstrap distributions of correlations between predicted and actual WM scores showed no significant difference in predictive power between models trained on movie-watching and n-back data (e). The bootstrap distribution revealed that models trained on movie-watching and n-back data exhibited higher correlations than those trained on resting state data (e). The Grad-CAM-derived saliency map highlights dominant features in the FC maps that contributed to the model’s predictions (f). The hot spots overlaid on an FC map demonstrate noticeable cross-correlation contributions in the “VAN”, “visual”, and to a lesser degree (<.5) “DMN”. Other important features visualized by Grad-CAM include off-diagonal hot spots reflecting inter-connections of the “DMN” – with “FPN, Fronto-parietal Task Control”, “Subcortical”, and “Cerebellar”; “Cerebellar” – “FPN” node.
The boxplots compare physical activity scores (total hours per week), CVD risk, and educational level (years) between two groups with positive and negative brain-cognition gaps in the DyNAMiC (top row) and COBRA (bottom row) datasets. The striped cyan-colored boxes represent scores for subjects with a positive brain-cognition gap, while the magenta-colored boxes represent subjects with a negative estimated gap.
Relationship between the gap measured from predicted and actual EM scores and dopamine D1 receptor (a). Negative gaps indicate that the predicted EM score was lower than actual EM scores, while the positive gaps indicate higher predicted scores than actual EM scores. Partial correlation analysis showed a significant correlation between D1 receptor values and the measured negative and positive gaps. Relationship between the gap measured from predicted and actual EM scores and dopamine D2 receptor (b). Negative gaps indicate that the predicted EM score was lower than actual EM scores, while positive gaps indicate higher predicted scores than actual EM scores. Partial correlation analysis showed a significant link between D2 receptor values to negative gaps and positive gaps.
Overview of the experimental procedure and the use of datasets. We used a 3-fold within-sample (DyNAMiC) cross-validation where we trained our model on 120 subjects (8:2; 80% training:20% validation during training) and tested it in a separate sample of 60 subjects. The winning within-sample model was used for between-sample (COBRA) cross-validation.
(a) Example of functional connectivity map across three different cognitive states. SEN hand: SENsory hand; SEN Mouth: SENsory Mouth; CON: Cingulo-Operculum control Network; Aud: Auditory; DMN: Default Mode Network; Memory Ret: Memory Retrieval network; Visual: Visual; FPN: Fronto-Parietal Network; Salience: Salience control network; Subcortical (upper row): subcortical network included in original Power parcellation; VAN: Ventral Attention Network; DAN: Dorsal Attention Network; Cerebellar: Cerebellar network; Subcortical (lower row): additional Subcortical regions, including hippocampus and caudate, added to the original Power Parcellation; Uncertain: Regions with less known network assignment. (b) DenselyAttention architecture. Enhanced Residual Block (ERB) and High-Frequency Attention Block (HFAB) into the Transition Block. Note that each “D.L.” layer in the table corresponds to the sequence BatchNormalization-ReLU-Conv3×3.
Model trained on functional connectivity maps acquired at rest predicts (a) episodic memory (EM) of the test dataset. The models trained on movie-watching (b) but not n-back (c) datasets predicted EM scores of the test dataset significantly. Table S1 summarizes the p values and correlation power for each model. Test datasets were obtained from the same cohort for rest, movie-watching, and n-back. Bootstrap distributions of correlations between predicted and actual EM scores indicated no significant difference in the predictive power of EM between models trained on resting-state and movie-watching datasets (d). Additionally, the bootstrap distribution revealed that models trained on resting-state and movie-watching data yielded higher correlations than those trained on n-back data (d).
Correlation results for episodic memory (EM) score predictions – DyNAMiC dataset with Schaefer300 parcellation.
Model trained on functional connectivity maps acquired at rest did not significantly predict (a) the working memory (WM) of the test dataset. The corresponding model best performed in WM scores predictions while trained on the movie-watching data set (b). The model trained on the n-back dataset (c) was the second-best, predicting WM scores. Table S2 summarizes the p values and correlation power for each model. Test datasets for resting-state, movie-watching, and n-back tasks were derived from the same cohort. Bootstrap distributions of correlations between predicted and actual WM scores showed no significant difference in predictive power between models trained on movie-watching and n-back data (d). Furthermore, the bootstrap distribution revealed that models trained on movie-watching and n-back data exhibited higher correlations than those trained on the resting state dataset (d).
