Towards personalized precision functional mapping in infancy

Abstract

The precise network topology of functional brain systems is highly specific to individuals and undergoes dramatic changes during critical periods of development. Large amounts of high-quality resting state data are required to investigate these individual differences, but are difficult to obtain in early infancy. Using the template matching method, we generated a set of infant network templates to use as priors for individualized functional resting-state network mapping in two independent neonatal datasets with extended acquisition of resting-state functional MRI (fMRI) data. We show that template matching detects all major adult resting-state networks in individual infants and that the topology of these resting-state network maps is individual-specific. Interestingly, there was no plateau in within-subject network map similarity with up to 25 minutes of resting-state data, suggesting that the amount and/or quality of infant data required to achieve stable or high-precision network maps is higher than adults. These findings are a critical step towards personalized precision functional brain mapping in infants, which opens new avenues for clinical applicability of resting-state fMRI and potential for robust prediction of how early functional connectivity patterns relate to subsequent behavioral phenotypes and health outcomes.

Keywords: brain development; infants; precision network mapping; resting-state fMRI; resting-state functional brain networks; template matching.

Fig. 1.

Split-halves reliability analysis. (A) Average minutes of resting-state data in split halves in infant subjects (note that the number of minutes contained in each half is not necessarily equal due to variation in the number of frames included after motion correction). (B) Two examples per cohort of individual network maps generated from the first and second halves of split resting-state time series using template matching. Normalized mutual information is used to compare the similarity of the networks generated from the first versus second half of each subject’s functional data and compared to a null distribution generated from comparing inter-subject network maps. Aud, auditory network; DMN, default mode network; PMN, parietal medial network; Smd, sensorimotor dorsal network; VAN, ventral attention network; CO, cingulo-opercular network; FP, frontal parietal network; PON, parietal occipital network; SMl, sensorimotor lateral network; VIS, the visual network; DAN, dorsal attention network; MTL, medial temporal network; Sal, the salience network; Tpole, temporal pole network.

Fig. 2.

Comparison of select infant versus adolescent network templates. All major adult networks were observed in the infant template. The color bar displays the strength and valence of functional connectivity (FC) between a given seed region and the rest of the brain. Though the templates are initialized with the same network maps, the infant template differs from the adolescent template in several ways. One striking difference was the strong anticorrelation with the sensorimotor region present in infants, particularly evident in DAN, DMN, and FP (blue arrows). Other notable differences with these particular infant templates are the relatively strong anticorrelation between DAN and insula (fuchsia arrows) and between FP and the visual cortex (yellow arrows).

Fig. 3.

Normalized mutual information (NMI) used to quantify the similarity of network maps derived from split-halves reliability analysis and across methods (template matching vs. Infomap). (A) Using template matching, the similarity of within-subject network maps was significantly greater than inter-subject similarity for both adolescent and infant cohorts, indicating that template matching effectively resolves individualized network maps in infants. (B) Population statistics comparing the similarity of within- and inter-subject network maps across cohorts. Interestingly, infant network maps were generally more similar across individuals (i.e., inter-subject NMI was higher) compared to adolescents. (C) Contrary to template matching, Infomap-derived network maps did not show a significant difference between within- and inter-subject network similarity with split-halves analysis. (D) Direct comparison of network maps from TM versus Infomap derived from the full amount of resting-state data available for each infant (OHSU cohort only): same subject network maps compared between methods were significantly more similar than inter-subject network maps. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001.

Fig. 4.

Within-subject network map similarity is significantly associated with the number of minutes of rest available per subject. The plots show a significant positive linear correlation between the average minutes of rest per half for each subject and within-subject NMI in OHSU (left) and WashU (right) cohorts. Gray boxes are 5th to 95th percentiles of the distribution of within-subject network map similarity; dashed lines are group-average NMIs.

Fig. 5.

Similarity of within-subject network maps generated from various intervals of resting-state data. For each subject, network maps were generated from continuously (left plot) and randomly (right plot) sampled data in 1-, 5-, 10-, 15-, and 25-minute segments from the first half of the dense time series (averaged over 10 repetitions). The resulting network maps were compared to the network maps derived from the full second half of the time series. The blue boxes display the 5th to 95th percentiles (0.43 and 0.59 NMI) of group-wide within-subject similarity values; the gray boxes display the same for group-wide inter-subject similarity (0.38 and 0.53 NMI); and the average NMIs are displayed as dashed lines. Each color-coded line in the plots represents data from an individual subject in the sample.