Predicting age from cortical structure across the lifespan

Abstract

Despite interindividual differences in cortical structure, cross-sectional and longitudinal studies have demonstrated a large degree of population-level consistency in age-related differences in brain morphology. This study assessed how accurately an individual's age could be predicted by estimates of cortical morphology, comparing a variety of structural measures, including thickness, gyrification and fractal dimensionality. Structural measures were calculated across up to seven different parcellation approaches, ranging from one region to 1000 regions. The age prediction framework was trained using morphological measures obtained from T1-weighted MRI volumes collected from multiple sites, yielding a training dataset of 1056 healthy adults, aged 18-97. Age predictions were calculated using a machine-learning approach that incorporated nonlinear differences over the lifespan. In two independent, held-out test samples, age predictions had a median error of 6-7 years. Age predictions were best when using a combination of cortical metrics, both thickness and fractal dimensionality. Overall, the results reveal that age-related differences in brain structure are systematic enough to enable reliable age prediction based on metrics of cortical morphology.

Keywords: aging; brain morphology; cortical complexity; fractal dimensionality; gyrification; structural MRI.

Figure 1

Inflated and pial surfaces and an oblique coronal slice, from a young adult (20-year-old male), illustrating the seven parcellation approaches used.

Figure 2

Cortical surfaces for each age decade, sex, and training sample. Representative cortical surface reconstructions for individuals with ages in the first year of each decade (with the exception of 40s, where there were insufficient male participants between ages 40 and 41). All reconstructions are shown at the same scale. 3D reconstruction images were generated as described in Madan (2015).

Figure 3

Overview of factors known to influence estimates of brain morphology.

Figure 4

Age distributions for each of the datasets. (A) Age distributions for each of the datasets used in the training. (B) Age distribution for the aggregated training dataset (i.e., combining IXI, OASIS, and DLBS). (C) Age distributions for the independent test datasets.

Figure 5

Illustration of how fractal dimensionality is measured from a 2D structure. Modified from Madan and Kensinger (2016).

Figure 6

Regional age-prediction differences (R2) using the (A) DKT, (B) von Economo-Koskinas atlases, and (C) Brainnetome for the respective structural measures using smoothing-spline regression.

Figure 7

Age-prediction performance for the models based on each of parcellation atlases and structural measures. ‘Regions’ corresponds to the number of regions/predictors used in the model (see Figure 1). ‘FD’ denotes fractal dimensionality of the filled structures. Each box-and-whisker bar denotes the median (i.e., MdAE) as a tick-mark. The box spans the 25th to 75th percentiles; the whiskers span the 10th to 90th percentiles. Values for each of these percentiles, as well as R2, are reported in Table 1.

Figure 8

Scatter plot of the actual and predicted age for the best-fitting model.