Lifespan Gyrification Trajectories of Human Brain in Healthy Individuals and Patients with Major Psychiatric Disorders

Cortical gyrification of the brain represents the folding characteristic of the cerebral cortex. How the brain cortical gyrification changes from childhood to old age in healthy human subjects is still unclear. Additionally, studies have shown regional gyrification alterations in patients with major psychiatric disorders, such as major depressive disorder (MDD), bipolar disorder (BD), and schizophrenia (SCZ). However, whether the lifespan trajectory of gyrification over the brain is altered in patients diagnosed with major psychiatric disorders is still unknown. In this study, we investigated the trajectories of gyrification in three independent cohorts based on structural brain images of 881 subjects from age 4 to 83. We discovered that the trajectory of gyrification during normal development and aging was not linear and could be modeled with a logarithmic function. We also found that the gyrification trajectories of patients with MDD, BD and SCZ were deviated from the healthy one during adulthood, indicating altered aging in the brain of these patients.

To confirm the GI trajectories of patients with psychiatric disorders are different from the GI trajectory of healthy subjects, we estimated the parameters of Eq. 1 for each diagnosis group using resampling technique within each group. The fitting parameters, a and b, were not different across the HC in the three cohorts (Fig. S4). We found that the parameter a was lower than HC in MDD (Cohen's d = −1.68; p < 0.0001) and higher in BD-I (Cohen's d = 0.81; p < 0.0001) and SCZ (Cohen's d = 1.22; p < 0.0001), while the parameter b, which indicated the decrease rate of GI, was higher (less negative) than HC in MDD (Cohen's d = 1.66; p < 0.0001) and lower (more negative) in BD-I (Cohen's d = −1.35; p < 0.0001) and SCZ (Cohen's d = −1.48; p < 0.0001) than HC (Fig. 3C). These results showed that the GI of BD-I and SCZ decreased faster than HC during aging.
The gyrification indices in multiple brain regions were significantly lower in the patients with BD-I and SCZ than HC during aging after the age of 40 (Fig. S6). No significant difference was found in MDD patients. The group-level comparison was performed at each vertex of the cortical surface. Only vertices with p values lower than the threshold p values obtained at a false discovery rate (FDR) 26, 27 of 0.05 were shown. No further clustering adjustment was performed. To confirm the validity and uniformity of the combined HC sample, the HC of each of the three cohorts was compared to the combined HC sample, and no vertex was significantly different (Fig. S5). We also compared each of the patient samples to the corresponding HC sample within each cohort instead of the combined HC sample, and the results were similar to but less sensitive than the results with all HC (see Supplementary Materials Fig. S7).

Discussion
To our best knowledge, this is the first in vivo study to show the cortical gyrification index trajectory of the human brain over the lifespan, from age 4 to 83. During this age range, the GI trajectory follows a logarithmic function Note that the thin light gray area along the black line indicates the 95% fitting confidence of the trajectory. (B) Brain gyrification changes over lifespan. Regional gyrification was high in childhood and adolescence, but the decrease rate was also high during these periods. Overall gyrification indices across the brain were low in adulthood, but the decrease rate was also low. The age ranges were chosen to best represent the gyrification changes. Only lateral view is shown, because the gyrification changes of medial regions over the lifespan was much less than the lateral regions (see Fig. S3). The colors represent the GI values. of age. This is an important advancement of our knowledge on human brain development and aging, as previous studies showed a linear decrease of GI during the adulthood 4, 28 . Based on post-mortem studies, the gyrification increases after birth 29,30 , which was confirmed by a recent longitudinal imaging study on children before the age of two 31 . However, when the gyrification of human brain reaches its peak during development is still unknown. Although there have been studies suggesting that the brain complexity keeps increasing until teenage 32 and certain brain regions such as entorhinal cortex may have increased thickness until the age of 30 33 , our study suggests that the GI starts to decrease already around the age of four. The cellular mechanisms underlying the gyrification are still unclear 34 , although theories have been raised based on mechanical tension 35 , stress-dependent folding and differential growth 36,37 , regulated radial and tangential expansion 38 , axonal pushing 39 and minimization of effective free energy associated with cortical shape 3,30 . Most of these theories focused on the brain development during the young age, and the trajectory observed in the present study may be associated with different mechanisms. Thus, the gyrification trajectory established in this study will be valuable information that needs to be considered in the future gyrification theories considering both brain development and aging.
Our findings revealed abnormal gyrification trajectories in patients with major psychiatric disorders. The GI in patients with BD-I and SCZ decreased faster than HC during aging. The brain regions with altered gyrification indices in BD-I and SCZ are consistent with several previous brain imaging studies investigating them individually 19,20,40,41 . Patients with SCZ showed a significant decrease of GI in the dorsolateral prefrontal cortex, anterior cingulate cortex and supra-marginal cortex. The prefrontal cortex is known to be crucial for executive function 42 and working memory 43 . These important cognitive functions are impaired in SCZ 44 , which has been linked to the dysfunction of dorsolateral prefrontal cortex, such as low activity, low N-acetylaspartate concentration and abnormal dopamine metabolisms [45][46][47] . The anterior cingulate cortex is the key region for conflict monitoring, motivation modulation and mood regulation [48][49][50] , the pathology of which has been shown in mood disorders and schizophrenia [51][52][53][54] . The abnormally decreased gyrification indices of these regions are in line with the altered cognitive monitoring and executive control 55 in SCZ. The GI decrease in patients with BD-I was less extensive compared to patients with SCZ. However, lower GI in BD-I than HC in inferior frontal regions was consistent with previous studies on cortical gray matters 56 . We also found that the gyrification trajectory of patients with MDD might be different from HC, as well as from BD-I and SCZ, but no brain region showed significant alteration of gyrification in MDD. In fact, several studies with small sample size showed higher gyrification in MDD than in healthy controls 57,58 , while some showed lower 18,59 , which could be due to differences in genes or brain connections 57,59 . Further studies will be necessary to clarify these inconsistent results. However, our results did show a possible decrease of gyrification index in MDD at a young age, which can be confirmed with future longitudinal studies focusing on pediatric MDD. The more significant decrease of gyrification in BD and SCZ compared to MDD could also be related to the possible damages due to manic episodes in BD and positive symptoms in SCZ indicated in the neuroprogression theory of BD and SCZ 60,61 . In addition, this is in line with recent studies, which suggested that the number of manic episodes seems to be the clinical marker more robustly associated with brain changes and neuroprogression in BD 61 . Given that the psychiatric disorders have been a major burden 62 in this aging world 63 , brain markers of abnormal aging, such as the gyrification index, may help us understand the mechanisms of brain aging, evaluate future strategies to slow down the degenerative process and relieve the burden caused by these major psychiatric disorders.
Some limitations should be considered in our study. The GI trajectory over the lifespan was developed using only cross-sectional data, and longitudinal data with multiple follow-ups were necessary to map the individual GI trajectory and to compare the individual differences of the GI trajectory during aging. The GI obtained with the automated algorithm was slightly higher than the ones in the previous studies 31, 64 . However, a recent study showed a similar range of GI using the same automated algorithm 65 . Our study covered a large age range of human lifespan from 4 to 83, but the GI trajectory before the age of four was not included. This was due to multiple limitations, e.g., the data that were publicly available over the lifespan usually did not include children below the age of four; there were challenges to perform MRI scans on young children due to the head motion; and validation of the current automated algorithm on young children was still necessary.
In summary, the present study demonstrated for the first time that the gyrification of the human brain in vivo decreased non-linearly from the age of 4 to 83 and this process could be modeled with a logarithmic function of age. The results were consistent across three independent cohorts. Moreover, by comparing the gyrification trajectories of healthy subjects and patients diagnosed with major psychiatric disorders, such as major depressive disorder, bipolar disorder, and schizophrenia, we provided consistent evidence that these disorders were associated with abnormal gyrification during aging. These results will advance our knowledge about how our brain changes during normal and abnormal aging.

Materials and Methods
Samples. Structural brain images from 881 subjects were collected with T1-weighted magnetic resonance imaging (MRI) from three independent cohorts: a sample collected at San Antonio (the SA sample), a sample collected at Nathan Kline Institute -Rockland (the NKI sample) 66 and a sample collected by the Centers of Biomedical Research Excellence (the COBRE sample). The latter two samples were publicly available. All patients completed the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID) and met the corresponding DSM-IV criteria for MDD, bipolar 1 disorder (BD-I) or SCZ. Healthy controls (HC) had no current condition or history of psychiatric dysfunctions. All subjects had no history of neurological diseases. Signed informed consent have been obtained from all the subjects. No identifiable image from a specific subject was used for publication. The study protocols were approved by the Institutional Review Boards at the University of Texas at San Antonio in accordance with their guidelines and regulations.
The structural brain images were preprocessed and the cortical surface of each brain was reconstructed with Freesurfer 67, 68 (version 5.3, http://surfer.nmr.mgh.harvard.edu/). The reconstructed images were visually inspected by the authors to exclude the apparent reconstruction errors. The local gyrification index (GI) at each vertex of the reconstructed cortical surface mesh was calculated using the toolbox in Freesurfer with default settings 69 . Briefly, a circular region of interest was delineated on the outer surface, and its corresponding region of interest on the inner cortical surface was identified using a matching algorithm as described elsewhere 70 and the ratio between the folded inner cortical surface and its corresponding exposed outer surface was calculated as GI. The resulted GI values of each subject were smoothly sampled (Gaussian kernel, 10 mm) onto an average template provided by Freesurfer (fsaverage; 163842 vertices), so that cross individual comparison could be performed. The averaged GI for each subject was also calculated to represent the gyrification of whole brain.
The effect of brain volume on GI was taken into account for by adjusting the GI values with a linear regression of intracranial volume (ICV) on GI at each vertex and the whole brain GI, because previous studies found that the GI could be correlated with the brain volume 31, 71 . Modeling Gyrification Trajectory as a function of Age. We compared the fitting of the whole brain GI with age from ten common mathematical functions and chose the best function based on the averaged mean squared errors (MSE). The MSE of each fitting function was calculated with the non-linear fitting functions of the Statistics and Machine Learning Toolbox in MATLAB (The MathWorks, Inc., Natick, Massachusetts, United States) using default settings. Because the COBRE sample only included adults, it was not involved in the function selection process. For SA and NKI samples, the MSE of each function was calculated based on the residuals derived from two types of validation procedures: (a) a leave-one-out cross-validation within each of the two samples and (b) a cross-sample validation, in which a fitting function was optimized for one sample and tested on the other. Thus, we had four MSEs for each function and the averaged of these MSEs was used as the fitting performance for the corresponding function as shown in Table S1. The best function was the logarithmic function with three parameters.
As a result, we used the following logarithmic function of age to fit the average GI trajectory: where a was an indicator of GI levels independent of age, b controlled the decrease rate of GI over age and c was the translational term of age. The initial value of a was set to 4, because the mean of the brain average GI values of all the HC subjects was around 3 and the maximum whole brain GI was near 4. The initial value of b was set to −1, because GI apparently decreased with age. The initial value of c was set to 0. We found that c was sensitive to the age range of the sample and we had varied age range in COBRE HC, MDD, BD and SCZ samples. Furthermore, c did not provide any information of the GI levels and trajectory shapes, and its value was consistent in SA and NKI HC samples. In order to reliably compare the GI trajectories of all the samples without changing the shape and levels of GI trajectory, we fixed the c value to −2.9991 according to the cross-validated fitting for the combined SA and NKI HC sample. Then each of the MDD, BD and SCZ groups was fitted separately. GI was adjusted for study site by an amount of the estimated GI difference between the HC in each cohort and the combined SA and NKI HC sample at a given age [a HC + b HC * ln(age + c)] − [a X + b X * ln(age + c)], where X represents the sample of HC in SA, NKI and COBRE cohorts and the mean adjustment values were as small as −0.0023, 0.0043 and 0.0523, respectively. The adjustment for the COBRE cohort was larger than the other two cohorts, because the GI in the COBRE HC sample was generally lower than the HC in the other two cohorts, which could also be observed in the mean GI of adult HC in the COBRE cohort (2.748) compared to the GI of adult HC in SA (2.7938) and NKI cohorts (2.7777). To further confirm whether the adjustment introduced significant effect in the SCZ sample in the COBRE cohort, we also compared the GI over the brain of HC and SCZ within the COBRE cohorts.
Although the samples showed different distributions of males and females (see also 72 ), which might have an effect on the GI trajectory of the samples with major psychiatric disorders, we found that effect of gender on GI was negligible after GI was adjusted by the ICV (Supplementary Materials Fig. S1). This might indicate that the gender effect on GI could be mostly explained by the brain volume. Thus, no further adjustment for gender was performed.
Estimating the Parameters with Resampling. In order to quantify the difference between the GI trajectories of HC, MDD, BD-I and SCZ, it was necessary to estimate the distributions of fitting parameters. We utilized the resampling technique. Briefly, for each of the HC, MDD, BD and SCZ samples, the sample was re-sampled 10000 times and the fitting was performed for each of the 10000 samples. In order to stratify the age distribution during the resampling, each sample was divided into four age blocks: 4-9, 9-18, 18-40 and 40-83. During each iteration of the resampling, 50% of the subjects in each age block were selected without duplicate. Thus, we had 10000 sets of parameters for each sample, and it was then possible to estimate the distributions of the fitting parameters in different samples.