Fig 1.
Method overview: step1) training semi-supervised AE model with joint optimization of age and sex prediction. Step2) applying UMAP transformation to the latent variables of semi-supervised AE. step3) applying HBR normative modeling to the components of UMAPs. Step4) measuring the correlation of non-imaging scores (behavioral, cognitive and clinical scores ) and the deviation value from normative range of UMAP components (latent representation index).
Table 1.
Model performance.
Fig 2.
A) UMAP representation of the latent space of selected contrasts from the Human Connectome Project (HCP) dataset according to Barch 2013, colored to show age and sex separation. B) UMAP representation of the latent space of HCP task contrasts, showing task separation. This is identical to panel A, except that the data points are colored according to task instead of age and sex C) UMAP representation of the latent space of the Faces-Shapes task contrast in the UKB dataset.
Fig 3.
A) Generated Motor-task contrast: Changes in the input space activation correspond to moving through the centroid of the latent space for HCP Motor Control subtasks. B) Generated HCP contrast: Projections of the centers of the latent representation contrasts (according to Barch 2013) into the input image space. C) Generated UKB contrast: Projections of the centers of the latent representation of the Faces-Shapes subtask into the input image space. Abbreviations: LF: left foot; LH: left hand; RF: right foot; RH: right hand; T: tongue.
Table 2.
Correlation of activations at the center of latent in the original image space with previous findings, as derived from Neurosynth meta-analytic database.
Each row represents a specific reported task in existing literature, along with the corresponding correlation values for each subtask in HCP and UKB datasets. Note that the corresponding subtasks per task are: Emotion (Faces-Shapes), Gambling (Reward-Punish), Language (Story-Math), Motor (AVG), Relational (REL), Social (Theory of Mind), and Working Memory (2BK-0BK).
Fig 4.
Normative models of the latent space UMAP components for males and females.
The figure presents four separate normative models for two UMAP components of the latent space, each depicting the relationship between age and the corresponding UMAP component. The individualized deviations from the normative range represent the latent representation index. Each percentile line within each figure displays the level of deviation from the normative range for each time point, illustrating the degree to which individuals differ from the expected normative pattern across various percentiles.
Fig 5.
Manhattan plot of p-value of univariate correlation of non-imaging measures with the individualized deviations from normative UMAPs of latent space (latent representation index) and PCA.
The black line is Bonferroni-corrected p-value threshold.