Mapping human brain charts cross-sectionally and longitudinally

Abstract

Brain scans acquired across large, age-diverse cohorts have facilitated recent progress in establishing normative brain aging charts. Here, we ask the critical question of whether cross-sectional estimates of age-related brain trajectories resemble those directly measured from longitudinal data. We show that age-related brain changes inferred from cross-sectionally mapped brain charts can substantially underestimate actual changes measured longitudinally. We further find that brain aging trajectories vary markedly between individuals and are difficult to predict with population-level age trends estimated cross-sectionally. Prediction errors relate modestly to neuroimaging confounds and lifestyle factors. Our findings provide explicit evidence for the importance of longitudinal measurements in ascertaining brain development and aging trajectories.

Keywords: brain trajectory; cross-sectional; individual prediction; longitudinal; normative models.

Fig. 1.

Study design. (A) GAMLSS frameworks established normative reference ranges of variation in cross-sectional baseline MRI measures (GMV, CTh, SA, and FA). (B) Numerical differentiation of the fitted cross-sectionally derived normative curves yielded group-level rates of change in MRI measures (red line), which were compared to rates of change directly ascertained from longitudinal measurements (blue line). Group-level estimates of rates of change were used to predict individual phenotype measurements at follow-up in unseen subjects. Specifically, individual trajectories were predicted from i) naive models assuming no change over time (i.e., equivalent baseline and follow-up values); ii) rates of change derived from cross-sectional data (50-th percentile); and rates of change derived from cross-sectional data using each individual’s percentile at baseline (SI Appendix, Fig. S5). (C) Errors in predicting individualized brain trajectories (i.e., mean absolute errors) were examined for (D) associations with demographic characteristics, neuroimaging confounds, and lifestyle variables across four domains: alcohol consumption, physical activity, sleep, and tobacco smoking. Abbreviations: gray matter volume (GMV); cortical thickness (CTh); surface area (SA); fractional anisotropy (FA); Cross (cross-sectional).

Fig. 2.

Normative models of brain aging. Normative centile reference ranges for (A) cross-sectionally measured whole-brain GMV, CTh, SA, and FA and (B) rates of change for each phenotype, shown alongside 25% and 75% CIs (dashed lines), generated with bootstrapping (500 samples). Rates of change were estimated i) directly from longitudinally measured phenotypes (blue) and ii) by differentiating the median centile curves (red). Insets show the percentage difference between cross-sectional and longitudinal rates of change. Abbreviations: gray matter volume (GMV); cortical thickness (CTh); surface area (SA); fractional anisotropy (FA); diff (difference).

Fig. 3.

Predicting individualized trajectories from group-level cross-sectional trends. (A) Observed change from baseline to follow-up (X axis) and predicted rate of change, as estimated from cross-sectional (cross) baseline data (using the 50-th for all individuals). (B) Mean absolute error in predicting the rate of change with cross-sectional models based on the 50-th percentile (mean) and individualized percentiles at baseline (indv %) and naive models (i.e., follow-up phenotype values are equal to baseline phenotype values). Bars denote between-group comparisons and asterisks denote significance (P_FDR < 0.01). Predictions derived from median centiles and individual-specific centiles yielded comparable accuracy for all phenotypes (no significant differences, P > 0.05). Abbreviations: gray matter volume (GMV); cortical thickness (CTh); surface area (SA); fractional anisotropy (FA); Cross (cross-sectional).

Fig. 4.

Neuroimaging and lifestyle associations with errors in predicting individualized rates of change. (A) MAE variance explained (r²) by neuroimaging factors, including PCA-summarized neuroimaging confounds measured at baseline (derivation of PCA components are shown in SI Appendix, Fig. S7), direction of change (increasing/decreasing), and interscan time interval (B) Regression plots present significant correlations between lifestyle measures (Y axis) and normalized MAE (X axis) from cross-sectional predications of change, based on the 50-th percentile (X axis). Asterisks denote significance (P_FDR < 0.05). Abbreviations: Principal component analysis (PCA); mean absolute error (MAE); cross-sectional (cross); individual (indv).