If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Most psychiatric disorders emerge in the second decade of life. In the present study, we examined whether environmental adversity, developmental antecedents, major depressive disorder, and functional impairment correlate with deviation from normative brain development in adolescence.
We trained a brain age prediction model using 189 structural magnetic resonance imaging brain features in 1299 typically developing adolescents (age range 9–19 years, mean = 13.5, SD = 3.04), validated the model in a holdout set of 322 adolescents (mean = 13.5, SD = 3.07), and used it to predict age in an independent risk-enriched cohort of 150 adolescents (mean = 13.6, SD = 2.82). We tested associations between the brain age gap and adversity, early antecedents, depression, and functional impairment.
We accurately predicted chronological age in typically developing adolescents (mean absolute error = 1.53 years). The model generalized to the validation set (mean absolute error = 1.55 years, 1.98 bias adjusted) and to the independent at-risk sample (mean absolute error = 1.49 years, 1.86 bias adjusted). The brain age estimate was reliable in repeated scans (intraclass correlation = 0.94). Experience of environmental adversity (β = 0.18; 95% CI, 0.04 to 0.31; p = .02), diagnosis of major depressive disorder (β = 0.61; 95% CI, 0.23 to 0.99; p = .01), and functional impairment (β = 0.16; 95% CI, 0.05 to 0.27; p = .01) were associated with a positive brain age gap.
Risk factors, diagnosis, and impact of mental illness are associated with an older-appearing brain during development.
). The highest risk of mental illness onset coincides with a period of rapid brain development. Cortical gray matter volume shows a nonlinear decline during adolescence, driven by reductions in cortical thickness and surface area (
Advances in machine learning (ML) algorithms combined with the availability of large open-access databases of brain scans have made it possible to predict a person’s age from their magnetic resonance imaging (MRI) data (
). By analyzing scans from typically developing individuals, ML determines the relationship between neuroanatomy and chronological age, across the age range, thereby mapping normative developmental trajectories. The ML algorithm is then able to make brain age predictions on new, previously unseen data. The difference between a person’s predicted brain age and their chronological age results in the brain age gap (
) disorders. A recent analysis of over 40,000 scans has revealed that clinical groups, including those with major depressive disorder (MDD), bipolar disorder, and schizophrenia, exhibited a positive brain age gap when compared with healthy control subjects (
). While the brain age concept is well established in adults, less is known about how the brain age gap relates to mental illness in adolescence.
Prior work has established the validity of brain age prediction in the developmental context, with a reliable differentiation between age groups (childhood, early adolescence, middle adolescence, late adolescence) (
). Building on the prior literature, we wanted to examine whether the brain age gap may work as a general biomarker of deviation from typical development. To do that, we examined the association between the brain age gap and multiple adverse exposures and outcomes.
In the present study we created a model optimized for predicting brain age in typically developing adolescents. We then applied the model in a high-risk sample of adolescents to calculate the brain age gap for each individual. Considering that there is no single indicator of ill mental health in the developmental context, we implemented a multiprong approach (
). First, we examined the association between multiple negative environmental factors, such as low socioeconomic status and maltreatment, and the brain age gap. Second, we investigated the association between early transdiagnostic symptoms for mental illness and the brain age gap. Third, we examined whether presence of MDD was associated with a wider brain age gap. Finally, we tested the relationship between functional impairment and the brain age gap. Given the prior evidence, we expected that all indicators would reflect in a positive brain age gap.
Methods and Materials
The current study includes data collected through data sharing platforms as well as an in-house cohort (Families Overcoming Risks and Building Opportunities for Well-being [FORBOW]). Table 1 provides detailed characteristics of each cohort. The study protocol was approved by the Research Ethics Board at Nova Scotia Health. Each cohort collected data with the participants’ informed consent and approval by local institutional review and ethics boards. Inclusion criteria for all cohorts included participants ages 9–19 years, passing automated MRI quality control. Figure 1B shows sample size, age, and sex distribution of the included cohorts. To model normative development, the training and validation sets (Figure 1C) only included control participants free from psychiatric disorders, with IQ above 75, where provided. An independent cohort (FORBOW), enriched for risk for mental illness, was used for hypothesis testing. FORBOW inclusion criterion was age 9–19 to capture development during adolescence. FORBOW exclusion criteria were 1) personal history of psychotic illness or autism spectrum disorder, 2) any serious medical or neurologic disorders, 3) MRI contraindications, and 4) full-scale intelligence quotient below 75 on the Wechsler Abbreviated Scale of Intelligence (
CoRR was created as an open science resource focused on test-retest reliability of MRI connectivity work. The data, including structural images, were gathered across the world and shared through the International Neuroimaging Data-sharing Initiative.
The NIH pediatric study utilized a uniform acquisition protocol and epidemiological sampling to form an MRI database of typically developing children. The data were shared with researchers and the clinical medicine community in the mid-2000s.
ABIDE, Autism Brain Imaging Data Exchange; CMI, Child Mind Institute: Healthy Brain Network; CoRR, Consortium for Reliability and Reproducibility; FORBOW, Families Overcoming Risks and Building Opportunities for Well-being; MRI, magnetic resonance imaging; NIH, The NIH MRI Study of Normal Brain Development; PING, Pediatric Imaging, Neurocognition, and Genetics.
The adversity score was calculated as a mean of 10 binary indicators of socioeconomic adversity and victimization: 1) biological mother’s education, 2) biological father’s education, 3) home-ownership status, 4) annual household income, 5) emotional abuse, 6) physical abuse, 7) sexual abuse, 8) neglect, 9) exposure to violence at home, and 10) bullying. Our previous work goes into greater detail on these indicators (
). In instances where binarizing for visualization was necessary, polyenviromic risk score (PolyE) was considered high if there was any childhood maltreatment.
Youth Experience Tracker Instrument
The Youth Experience Tracker Instrument (YETI) is a 26-item self-report measure of 6 psychopathological antecedents that precede and predict mental illness: affective lability, anxiety, depressive symptoms, basic symptoms, psychotic-like experiences, and sleep (
). The questionnaire was designed to facilitate early identification of risk for SMI and has been validated in the FORBOW cohort. A score ≥ 8 maximized the product of sensitivity and specificity when screening for early antecedents. The measure is available at https://www.youthhealthmeasures.com/.
Major Depressive Disorder
Participating youth were interviewed using the Kiddie Schedule for Affective Disorders and Schizophrenia, Present and Lifetime Version (
). The questionnaire taps into several functional domains, including relationships with family at home, relations with peers, academic functioning, and involvement in general hobbies/interests. A score ≥ 15 indicates clinically significant functional impairment (
). We took the average of the child and parent reporting on the CIS for the analysis.
Images were acquired with a 3T General Electric Discovery MR750 scanner (GE Healthcare) equipped with a 32-channel radiofrequency head coil (MR Instruments, Inc.). Scanning took place at the Biomedical Translational Imaging Centre, Halifax, Nova Scotia. We collected a three-dimensional T1-weighted brain volume imaging sequence with whole-brain coverage, 1-mm3 isotropic resolution, matrix = 224 × 224, field of view = 224 mm, 168 sagittal slices at 1-mm thickness, repetition time = 5.9 ms, echo time = 2.2 ms, inversion time = 450 ms, flip angle = 12°. In addition, we collected a T2-weighted Cube fluid attenuated inversion recovery sequence using a T2 prep contrast option with identical coverage, resolution, and acquisition orientation to the T1-weighted sequence; echo time = 98 ms, repetition time = 5100 ms, inversion time = 1427 ms, echo train length = 250 echoes, flip angle = 90°.
MRI Processing and Model Features
Data across all cohorts were processed with FreeSurfer (version 6.0.0) software; RRID:SCR_001847 (
). Qoala-T is an ML tool designed to automatically classify the quality of FreeSurfer output. Please see the Supplement for more details regarding the quality control. The brain age prediction model was trained only on data that were recommended for inclusion by Qoala-T.
For cortical features, we compiled 34 cortical structures per hemisphere based on the FreeSurfer default Desikan–Killiany atlas (
). We included cortical gray matter volume and cortical surface area measurements for each structure. We excluded cortical thickness measurements for several reasons: 1) cortical thickness measurements are the least reliable and most sensitive to MRI artifacts (
), and there is no agreed upon way to standardize them within a cross-validation framework; and 3) we wanted to maintain an adequate ratio of features/predictors (p) to sample size (n). Overall, we included 136 cortical features.
We also included 53 bilateral global (e.g., intracranial volume) and subcortical volumetric measures (e.g., amygdala volume) from the FreeSurfer output. The volume of the fifth ventricle was removed owing to near 0 variance. Overall, 189 features were used for the age prediction model (Table S3 for the full feature list).
Brain Age Prediction
First, we split the available data into training, validation, and test sets (Figure 1C). Eighty percent of the typically developing control subjects were used for model training. We implemented 10-fold cross-validation, repeated 10 times, with the folds stratified by the scan age. Twenty percent of the held-out control subjects were used for model validation and estimation of model bias (
); see the Supplement. We tested hypotheses in a separate at-risk cohort (FORBOW), fully independent from training and validation. The test set included 338 scans from 150 individuals; see Table S2 for the scan timing breakdown.
ML was performed within the tidymodels (version 0.1.1) framework (
). The regression models predicted scan age using cortical and subcortical brain features. The primary objective was to minimize the mean absolute error (MAE) metric, expressed in years. See Table S4 for additional model parameters. We used the finalized XGBoost model to predict scan age in the validation and test sets. The final model is available on GitHub. We computed the brain age gap by subtracting the true scan age (chronological age) from the predicted age (Figure 1A). We bias corrected the brain age estimation using prior methods (
). In order to report on the specific brain regions most relevant to predicting age, we created variable importance plots with the vip R package (version 0.2.2) and visualized those regions with the ggseg R package (version 1.5.4) (
), we covaried for the effects of age, age2, and sex in the main analyses. For sensitivity analysis, we additionally covaried family history of SMI, socioeconomic status, any prior cannabis use, and IQ. Socioeconomic status was not used as a covariate in the PolyE sensitivity analysis as it is one of the components of the adversity measure. To assess the relative impact of adversity indicators on individual development, we modeled and provided a breakdown of the subcomponents in the Supplement. We modeled multiple observations from each participant by including the individual identifier as a random effect. We also accounted for the nonindependence of data from related individuals by including the family identifier as a higher-order random effect. We reported effect sizes using standardized regression estimates (β) and their 95% confidence intervals (CIs).
Brain Age Model Performance
We built a brain age model using MRI scans of typically developing adolescents from 6 cohorts (n = 1299; age range 9–19 years old). The best model, selected for parsimony and accuracy, had an MAE of 1.53 years in cross-validation. See Table S5 for the top 10 models and their parameters. The finalized model from training performed with an MAE of 1.55 years on the held-out validation set (bias corrected MAE = 1.98). See Figure 2 for model performance across datasets. Our target of interest was the independent FORBOW cohort, where the finalized model performed with an MAE of 1.49 years (bias corrected MAE = 1.86).
Reliability of Predicted Age
While the reliability of individual brain imaging measurements has been established, the short-term reliability of brain age estimation is unknown. In 50 individuals re-scanned within weeks of their first scan, the intraclass correlation coefficient (1,1) for predicted brain age was 0.94 years (95% CI, 0.90 to 0.96). The reliability analysis shows that the data and the model are of sufficient quality to make reliable and consistent brain age predictions.
Neuroanatomical Contribution to Age Prediction
The ML model relied on a diverse set of features to make accurate age predictions. The top predictors included the total brain segmentation volume, the precuneus and other parietal regions, the bilateral ventral diencephalon, several prefrontal regions, and subcortical areas such as the hippocampus and the amygdala. Figure 3 displays the variable importance of the top cortical and subcortical volumes to brain age prediction.
Association With Exposures and Outcomes
First, we examined how an aggregate measure of adverse environmental exposures (PolyE) relates to brain development. After controlling for age, age2, and sex, we found that PolyE was significantly associated with a positive brain age gap (β = 0.18; p = .02; 95% CI, 0.04 to 0.31). In other words, adversity was associated with older-appearing brains (Figure 4A). This effect remained after additionally controlling for the effects of family history of SMI, any lifetime cannabis use, and IQ (β = 0.17; p = .02; 95% CI, 0.03 to 0.31). See Figure S1 and Table S6 for the associations between the individual subcomponents of the adversity score and the brain age gap.
Next, we tested the association between the YETI and the brain age gap. We found that the dimensional measure of 6 psychopathology antecedents was not significantly associated with deviation from chronological age (β = 0.10; p = .09; 95% CI, −0.01 to 0.21). Individuals high on the symptom scale were shifted to the right of the brain age gap distribution, toward an older-appearing brain; however, there was substantial overlap (Figure 4B).
Major Depressive Disorder
In Figure 4C, we showed that MDD was associated with a positive brain age gap (β = 0.61; p = .01; 95% CI, 0.23 to 0.99). The effect of an older-appearing brain in MDD held when further controlling for family history of SMI, socioeconomic status, any lifetime cannabis use, and IQ (β = 0.47; p = .02; 95% CI, 0.08 to 0.86). We wanted to explore the depression phenotype in the context of other factors that may be a risk for depression. See Figure S2 and Table S7 for the results from multivariate regressions modeling MDD along with the risk factors (adversity, antecedents) and outcomes (functional impairment).
Functional Impairment (CIS)
Finally, we examined the association between functional impairment and the brain age gap (Figure 4D). The main analysis showed that increasing functional impairment relates to a higher brain age gap (β = 0.16; p = .01; 95% CI, 0.05 to 0.27). The effect held in the sensitivity analysis (β = 0.13; p = .02; 95% CI, 0.02 to 0.24). See Figure S3 for a forest plot of associations across all the tests.
In this work, we were able to predict age using neuroanatomical data accurately and reliably in a large sample of typically developing adolescents. The prediction accuracy generalized to an independent cohort of adolescents enriched for risk of mental illness (FORBOW). Within FORBOW, we found that multiple factors were associated with discrepancies between the predicted and chronological age. The spectrum of adverse experiences and outcomes, such as environmental adversity, depression onset in adolescence, and functional impairment, were all associated with a positive brain age gap. However, the unique effects of each marker while accounting for the presence of others remains to be tested in larger samples. Deviation from chronological age might be a general indicator of unwellness rather than a marker of a specific exposure or outcome.
Our accuracy in predicting chronological age falls in line with prior developmental work with MAEs from 1 to 2 years [see (
) for comprehensive review of the last decade of brain age studies]. We were able to achieve a 1.5-year MAE while maintaining scale, generalizability, and reliability. To our knowledge, we have aggregated the largest developmental sample in the study of brain age and leveraged it to show consistent performance among training, validation, and independent testing.
The few published reports on developmental brain age suggest that an increased brain age gap is linked with psychopathology, including the risk of psychosis (
). The present study not just confirms the association between the brain age gap and depression but extends this to associations with known risk factors and functional impairment. Multiple types of adverse exposures in childhood and adolescence have been established as strong risk factors for depression and other type of illness (
). In line with our hypothesis, we found that cumulative exposure to adversity was associated with advanced brain age. This novel approach to capturing aggregate adversity can be interpreted in the context of existing neuroimaging literature on the underlying components of the dimensional measure. Community disadvantage and lower socioeconomic status have previously been associated with reduced cortical tissue volume, both globally and regionally (
). Our measure also assessed childhood maltreatment, including physical, sexual, and emotional abuse and neglect. Maltreatment has been shown to have lasting effects on the stress response system and neurobiological development (
The most prominent neuroanatomical associations with maltreatment include gray matter reductions in frontoparietal and limbic structures such as the precuneus, prefrontal cortex, and hippocampus, enlargements in the ventricles, and reductions in the size of the corpus callosum (
), the ML model is likely sensitive to the additional decreases associated with adversity, resulting in a prediction of advanced maturation or a wider positive brain age gap. Correspondingly, disproportionate increases in measures such as ventricular volumes, which typically increase with chronological age, are also associated with increased brain age. On average, the age prediction model was equally accurate on the training data from control individuals and the data from individuals at familial risk for mental illness. Yet, the brain age gap was increased in adolescents who experienced adversity. Our results suggest that in the interplay of biology and circumstance, deviation from chronological age may more strongly reflect what happens throughout life rather than the genetic contributions present from the beginning.
We also hypothesized that developmental antecedents for mental illness would be associated with an increased brain age gap. While higher levels of antecedents shifted the distribution toward an increased brain age gap, the effect was not statistically significant. This was unexpected because we previously found that the basic symptoms and psychotic-like experiences captured by the YETI were associated with reduced cortical folding in this cohort (
). It is possible that early preprodromal symptoms are associated with subtle cortical alterations that are masked amid the developmental changes occurring in adolescence. Our previous work found reductions in the orbitofrontal and occipital areas, none of which feature prominently in the list of structures that contribute most to accurate age prediction. Our current findings compare to those from Chung et al. (
), where prodrome symptoms were associated with a wider brain age gap but the trained models were not robust enough to predict conversion to psychosis.
In our final analyses, we examined the brain age gap in the context of depression and then concluded the investigation by demonstrating the functional relevance of the brain age gap beyond diagnostic group differences. In the present study, MDD was associated with a higher positive brain age gap, confirming our hypothesis and resembling the effect recently found in adults with MDD (
) has demonstrated that an increased brain age gap was associated with worse outcomes on several disorder specific functional measures. Our study expands on this finding by demonstrating a more generalized effect of functional impairment on the brain age gap during adolescence.
Limitations and Future Directions
There were several limitations to this study. Brain age prediction error can be further reduced in several ways. For example, some scan sites only provided chronological age as whole integers, imposing a built-in cap on prediction accuracy. Furthermore, state-of-the-art deep learning methods, specifically convolutional neural networks, have shown greater accuracy than more shallow ML techniques in adult samples (
). An additional benefit of convolutional neural networks would include fewer processing requirements for MRI data with traditional software. Furthermore, we focused on a broad range of cortical and subcortical measures. However, the feature set can be expanded to include multimodal data for improved accuracy and, perhaps more importantly, mechanistic inference (
The brain age gap addresses two major challenges and opportunities in neuroimaging: high dimensionality of brain data and individual-level prediction. Regions across the brain are analyzed in aggregate through a multivariate framework to make accurate individual-level predictions. This process is then summarized by an intuitive single number, the brain age gap. However, this reduction in complexity is a double-edged sword when it comes to identifying the mechanisms responsible for the brain age gap. Our efforts at interpretable ML showed which structures contributed most to brain age prediction. By incorporating diffusion-weighted imaging, intracortical myelination, and data from other modalities, future studies may be able to capture different aspects of pathophysiology contributing to brain age gaps. A great example of such work was done by Brown et al. (
). The group examined age-varying contributions of different imaging measures to the prediction of age in healthy individuals. For example, diffusion measures from white matter fiber tracts were found to be most relevant between the ages of 12–15 years. Improvements on such study designs, done at a population neuroscience level, might offer mechanistic insights into the pathophysiology of brain age deviation in the context of neuropsychiatric disorders.
Given the high covariance between adversity, depression, and impairment, a much larger sample would be required to conclusively identify the unique effects of each marker. A new generation of large-scale studies (
) investigating adolescent brain development across health and disease might offer enough power to examine individual markers while accounting for their correlation. Furthermore, longitudinal studies will be able to characterize the distinct developmental trajectories of at-risk individuals. Future work can run such analyses and target subsamples in further detail with adequate power to jointly examine the effects of lifestyle factors and medical conditions that are also known to influence brain health (
We have shown that several factors, including adversity, depression onset in adolescence, and functional impairment, were associated with a higher brain age gap. Future work in larger samples will allow us to study the independent contributions of these factors to brain age. We have shared our brain age algorithm with the wider scientific community. Future studies may characterize how an individual’s deviation from typical neuroanatomical maturity changes throughout life.
Acknowledgments and Disclosures
This study was supported by the Independent Investigator Award, Brain & Behavior Research Foundation (Grant No. 24684 [to RU]); the Canada Research Chairs Program (Grant No. 231397 [to RU]); the Canadian Institutes of Health Research (Grant Nos. 124976 , 142738 , 148394 , and 173592 [to RU]); Nova Scotia Health ; the Dalhousie Medical Research Foundation ; and a doctoral graduate award from the Canadian Institutes of Health Research (Grant No. 157975 [to VD]).
The Biomedical Translational Imaging Centre imaging facility has received funding support from Brain Canada.
VD, MHS, CVB, and RU designed the study. VD, HVG, CAH, and RU acquired the data, which VD and RU analyzed. VD wrote the article, which all authors reviewed. All authors approved the final version to be published and can certify that no other individuals not listed as authors have made substantial contributions to the paper.
We thank the participating families and acknowledge the contributions of the Families Overcoming Risks and Building Opportunities for Well-being research team (see http://www.forbow.org). We also thank Anna Nazarova and Anna Minarik for their help in the lab.