Atypical brain aging and its association with working memory performance in major depressive disorder

Abstract

Background:

Patients with major depressive disorder (MDD) can present with altered brain structure and deficits in cognitive function similar to aging. Yet, the interaction between age-related brain changes and brain development in MDD remains understudied. In a cohort of adolescents and adults with and without MDD, we assessed brain aging differences and associations through a newly developed tool quantifying normative neurodevelopmental trajectories.

Methods:

304 MDD participants and 236 non-depressed controls were recruited and scanned from three studies under the Canadian Biomarker Integration Network for Depression. Volumetric data were used to generate brain centile scores, which were examined for: a) differences in MDD relative to controls; b) differences in individuals with versus without severe childhood maltreatment; and c) correlations with depressive symptom severity, neurocognitive assessment domains, or escitalopram treatment response.

Results:

Brain centiles were significantly lower in the MDD group compared to controls. It was also significantly correlated with working memory in controls, but not the MDD group. No significant associations were observed in depression severity or antidepressant treatment response with brain centiles. Likewise, childhood maltreatment history did not significantly affect brain centiles.

Conclusions:

Consistent with prior work on machine learning models that predict “brain age”, brain centile scores differed in people diagnosed with MDD, and MDD was associated with differential relationships between centile scores and working memory. The results support the notion of atypical development and aging in MDD, with implications on neurocognitive deficits associated with aging-related cognitive function.

Introduction

Aging is associated with gradual physiological changes in brain and behavior. Age-related cognitive decline occurs in several domains, including memory, attention, and executive function(1,2). For example, a large-scale prospective cohort showed that memory, processing speed, executive function, and global cognition declined with older age(3). In turn, age-related cognitive decline correlates with global cerebral atrophy, as evidenced by reduced gray matter volume (GMV), cortical thinning, sulcal widening, and ventricular expansion on magnetic resonance imaging (MRI)(4). Using large databases, characterizations of normative brain development and aging have illuminated healthy aging brain at different ages(5). However, physiological age-related decline in cognition is heterogeneous(6); 1/19/2026 5:36:00 PMsome individuals decline faster than others of the same age, which may depend on environmental factors, genetics, or both(7).

Major depressive disorder (MDD) may significantly influence age-related decline(8). Individuals diagnosed with MDD exhibit sustained deficits in attention, working, and long-term memory, even after remission, and in greater effect in individuals with recurrent MDD(9). Furthermore, MDD patients present with altered brain structure and function like that observed in age-related cognitive decline, including gray matter atrophy regions crucial memory formation and processing, including the hippocampus, frontal cortex, putamen, thalamus, and amygdala(10). Unfortunately, relatively few studies focus on the relationship of aging and psychopathology. A comprehensive understanding of the pathophysiology of MDD is important for developing novel therapeutic strategies and optimizing existing ones. There exists a knowledge gap in understanding how depression impacts brain aging and cognitive function.

Brain-based age prediction is one method to understand the interplay between heterogeneous brain- and behavioral-based markers(11). Such approaches typically develop a machine learning model predicting age and trained on normative structural MRI data across the lifespan. Deviations between the chronological and predicted ages, also called the brain-predicted age difference (brain-PAD), can be applied to find differences between diagnostic groups or associations with behavior. There is considerable variability in the methodological approach (e.g., Gaussian process regression, regularized gradient tree boosting, deep learning), feature extraction (e.g., raw T1-weighted structural image, regional parcellation), and normative sample used for training. However, numerous studies have successfully investigated brain age using this method, including to predict mortality(11), multiple sclerosis progression(12), and dementia risk(13).

MDD patients may exhibit accelerated biological indices of aging. For example, MDD patients have shorter leukocyte telomere lengths relative to non-depressed controls (HC)(14,15) which may predict poorer response to selective serotonin reuptake inhibitors(16,17). Brain-PAD-based studies, however, present conflicting findings: some(18–20) but not all(21) demonstrated higher brain-PAD scores in MDD relative to HC. Similarly, brain-PAD-based characterizations of aging in MDD conflict with respect to pharmacotherapy response, with one group reporting no correlations between brain-PAD and escitalopram response(22) while another reported association between accelerated brain aging and poor sertraline response(23). Furthermore, these studies focused on adult cohorts and not in adolescents, which creates limitations for our understanding of risk for psychiatric disorders during critical neurodevelopmental windows.

The development of human brain charts addresses the need for a standardized tool to evaluate individual differences in age-related brain changes across the lifespan. Like height and weight growth charts, brain charts(5) present normative, non-linear trajectories of normative aging based on fitting Generalized Additive Models for Location, Scale and Shape models from a large (n>100,000), multisite structural MRI dataset including the UK Biobank, Human Connectome Project, and others. Global brain measures were used in BrainChart, including cortical grey matter, white matter, subcortical grey matter, and ventricular volumes. This approach was empirically optimized, was evaluated for confounds like site, and considered non-linear age-related changes in volume, separated by sex. Centile scores exhibit test-retest reliability in out-of-sample testing and robustness to varying image analysis pipelines. Although in its infancy, brain centile scores have potential clinical utility, showing significant differences in individuals with Alzheimer’s Disease and males with MDD(5). Brain charts could serve as a useful tool to investigate the relationship between aging and brain development. However, it is unknown whether brain centiles scores relate to cognitive factors related to aging, treatment response, or environmental factors conferring MDD risk(24).

Childhood maltreatment (CM) is one well-documented risk factor for MDD. CM correlates with earlier depression onset, greater severity, and a higher likelihood of developing treatment resistant MDD(25,26). Individuals with a history of CM also demonstrate structural changes in brain areas involved in emotional processing and memory, including the hippocampus and dorsomedial prefrontal cortex(27–30). CM interacts with age in estimating the cortical thickness of emotional regulation regions, like the insula, cingulate, orbitofrontal, dorsolateral, and medial prefrontal cortices(31). Therefore, investigating the effects of CM on brain age may elucidate its mechanism of increasing MDD risk.

This study aimed to examine the impact of brain aging—measured by brain chart-based centile scores—in MDD using a multi-site sample of individuals (12–65 years old). This study further aimed to investigate the impact of brain aging on age-related cognition, CM, and antidepressant response. We hypothesized that brain centile scores would significantly differ in individuals with MDD relative to non-depressed controls, with greater atypical centile scores correlating with depression severity. We also expected an association between centile scores and cognitive performance in age-related domains, and that this relationship would differ in MDD. Additionally, we predicted that brain centile score would be associated with response to the commonly prescribed first-line antidepressant medication, escitalopram. Lastly, we hypothesized centile scores would differ in individuals with and without a history of CM.

Methods

Recruitment

Healthy and depressed participants were recruited for three studies associated with the Canadian Biomarker Integration Network for Depression (CAN-BIND) program(32): Biomarkers of Antidepressant Response to Medication (CAN-BIND-1; NCT01655706), Canadian Psychiatric Risk and Outcome Study (PRO-CAN; NCT02739932), and Stress and Reward Anhedonia Study (SARA; NCT02798094). Details on the aims and design for each study are in the supplemental materials. We were adequately powered to assess differences in treatment response(33) and centile score. Using the results reported by Luo and colleagues(34) at alpha=0.05 and power=0.8, we could need 86 participants per group to detect a significant difference.

CAN-BIND-1 aimed to identified biological markers of pharmacotherapy response(33); recruitment occurred at six sites across Canada. All MDD participants (18–60 years old) were treated with 8 weeks of flexible-dose open-label escitalopram. HCs had the same age range and language requirements but no history of Axis I or II disorders. PRO-CAN sought to identify youth at risk of developing serious mental illnesses (SMI; 12–25 years old); this study recruited individuals with: (a) no mental health concerns, (b) an at-risk group with a family history of an SMI, (c) a group with early mood symptoms or (d) attenuated SMI symptoms(35). For our purposes, we retained HC and individuals with MDD symptoms meeting DSM-IV-TR criteria for a major depressive episode. SARA examined abnormalities in the processing of stressful and rewarding information and their relation to depression (18–65 years old). Both MDD and HC were recruited(36). Eligible participants of all studies provided written, informed consent and all study protocols were approved by Research Ethics Boards at each participating site.

Clinical Measures

CAN-BIND-1 and SARA measured depression severity using the Montgomery-Åsberg Depression Rating Scale (MADRS)(37), a 10-item, clinician-rated questionnaire. For CAN-BIND-1, the MADRS was acquired every 2 weeks throughout treatment. PRO-CAN assessed depression using the Beck Depression Inventory (BDI)(38), a 21-item, self-reported questionnaire. Consequently, we used MADRS and BDI as the primary depression severity measures. Values were combined by normalizing scale scores, generating z-scores.

CM was defined as a continuous measure including emotional, physical, and sexual abuse. For CAN-BIND-1 and SARA, CM history was collected using the Childhood Experience of Care and Abuse (CECA-Q)(39) scale, which coded emotional and physical abuse on a 4-point scale (Little/None; Some; Moderate; Marked) and sexual abuse on a 5-point scale (None; Little; Some; Moderate; Marked). PRO-CAN used a trauma documentation form to record trauma or abuse experienced before the age of 18(35), which included 5-point impact scale for all measures (None; Little; Moderate; Quite a Bit; Extreme). For this dataset, we adjusted the 5-point scale to a 4-point scale for emotional and physical abuse to align with CAN-BIND-1 and SARA by binning none and little selections).

Participants recruited in CAN-BIND-1 were treated with escitalopram. Antidepressant outcomes were quantified as both a percentage change in MADRS scores between baseline and Week 8, and a binary outcome of response (≥50% MADRS change).

Neurocognitive Measures

CAN-BIND-1 acquired the CNS Vital Signs (CNS-VS; RRID: SCR_024475), which is a tool that assesses ten neurocognitive domains: cognitive flexibility, executive function, composite memory, processing speed, reasoning, social cognition, sustained attention, visual memory, verbal memory, and working memory. For analysis purposes, we used percentile scores, which standardized an individual’s performance were extracted relative to an age-matched normative database.

Neuroimaging Acquisition and Preprocessing

The MRI protocols for all CAN-BIND studies have been previously reported(40). To summarize, all three studies obtained whole-brain T1-weighted structural scans with a 3-dimensional isotropic resolution of 1 mm. Structural neuroimaging data were acquired using 3.0 Tesla MRI systems, with various scanner models across sites; acquisition parameters are summarized in the supplemental materials.

As previously described(5), we preprocessed T1-weighted structural MRI scans using the standard recon-all pipeline in Freesurfer Version 7.1.0 (RRID: SCR_001847). Briefly, the first step of recon-all includes motion correction, non-uniform intensity normalization, projection to the Talairach space, skull stripping, and tissue/subcortical segmentation. Subsequently, the second and third steps serve to smooth, interpolate, and tesselate the data into surface space. We extracted the following tissue volume data for each participant from the “aseg.stats” file outputted by recon-all: total GMV; total cortical white matter volume (WMV); subcortical gray matter volume (sGMV), which encompassed the thalamus, caudate, putamen, pallidum, hippocampus, amygdala and nucleus accumbens; and ventricular volume (VV; the “volume of ventricles and choroid plexus” label).

Image quality was considered during acquisition and preprocessing. Participant instructions and support materials were uniform across all sites(40), and acquisition parameters were standardized where possible across all CAN-BIND sites. A participant also traveled to each scanning site to quantify inter-site variance. All scans were initially assessed by trained quality control raters, as it was being collected, for motion, field-of-view, or other artifacts. Participants were rescanned if necessary. After preprocessing, we executed quality control and reprocessed for improper segmentations if necessary, on 33 randomly selected scans, which represent approximately 5% of samples from different scanners, and scans with a value two standard deviations below or above the mean on any output.

Brain Centile Extraction

Participant demographic and clinical data, including age, sex, diagnosis, and MRI measures were compiled. The dataset was uploaded onto BrainChart (www.brainchart.io)(5) to obtain individualized centile scores that indicate the presence of any accelerated aging. Each centile score is computed by quantifying the vertical deviation of structural MRI phenotypes to the reference curves, which are stratified by sex. The tool incorporates an out-of-sample estimator of model parameters where maximum likelihood is used to estimate study-specific random effects; this allows the scoring of centiles using the cumulative density function.

Data Analysis

We used R-Studio V2022.07.0 (RRID: SCR_000432) to examine relationships between variables of interest. To account for scanner differences, we used the harmonization method ComBat on brain centiles and any MRI phenotypes used to generate it(41–43). We used generalized linear models (GLMs), and the Benjamini-Hochberg method(44) to control for False Discovery Rate (FDR). By characterizing brain aging with centile scores, the following questions were investigated:

1. Does brain centile score differ in participants diagnosed with depression relative to non-depressed controls? We conducted a GLM with diagnosis and ComBat-corrected centile score as the independent and dependent variables, respectively, and age and sex as covariates. Within the MDD group, including both single episode and recurrent MDD participants, we also used the GLM model to test whether the number of past depressive episodes, current episode duration, age of MDD onset, and depression severity predicts brain centile.

2. Which structural MRI measures drive differences in brain centile score by diagnosis? Analyses were performed separately for males and females; for each GLM model, diagnosis is the predictor variable, ComBat-corrected GMV, WMV, sGMV, or VV is the response, and age is a covariate. Results were corrected for multiple comparisons using FDR.

3. Do brain centile scores predict variability in neurocognitive domains associated with aging, and does this relationship differ between people diagnosed with MDD and non-depressed controls? In controls, we first predicted ten neurocognitive domains separately using ComBat-corrected brain centiles, with both sex and age as covariates. We then used brain centiles to predict ranked scores for domains that showed statistically significant results, while incorporating several covariates: sex, diagnosis, age, and interaction between brain centile and diagnosis. FDR multiple comparison correction was performed for results.

4. Does brain centile score differ in people with versus without a history of CM? Using a GLM, we predicted ComBat-corrected brain centile using overall CM and specific types of maltreatment; we added age, sex, diagnosis, and interaction between maltreatment and diagnosis as covariates. Multiple comparison correction was done using FDR.

5. Does brain centile score predict antidepressant response to escitalopram? In CAN-BIND-1 participants, a linear mixed effects model was used to further examine brain centiles with varying individual depression severity over 8 weeks. This model predicting MADRS, generated using the lmerTest package, included time, brain centile score, age and sex as fixed effects, and the intercept as a random effect. We used a time:centile score interaction to determine whether pre-treatment centile score was associated with escitalopram-related MADRS response. We also used a logistic GLM model with ComBat-corrected brain centile as the independent variable, antidepressant response as the dependent variable and included covariates such as age, sex, and baseline depression severity.

Results

Combining the CAN-BIND-1, PRO-CAN, and SARA datasets yielded a large, multi-site sample (Table S1) of moderately depressed patients diagnosed with MDD (n=304) and non-depressed controls (n=236). The MDD group was significantly older than HCs (Table 1). Sex did not differ by diagnostic criteria (Control, Single Episode MDD, Recurrent MDD; χ²=2.351, df=2, p=0.309). Depression severity was normally distributed in both single episode and recurrent MDD groups (Figure S1). We included age and sex as covariates in all subsequent models.

Table 1:

Descriptive statistics for demographic and clinical characteristics in control and major depressive disorder (MDD) groups for (A) all participants; (B) CAN-BIND-1; (C) PRO-CAN; (D) SARA.

	MDD			Control
	n	Mean	SD	n	Mean	SD	Test Statistic	p
A. All Participants
Age	304	32.806	12.958	236	27.987	11.820	3.696	2.42×10−4
Female/Male	202/102			142/94
Depression Severity	304
B. CAN-BIND-1
Age	192	34.75	12.553	107	32.850	10.483	1.726	0.086
Female/Male	125/67			69/38
MADRS	192	29.875	5.619
C. PRO-CAN
Age	20	18.15	2.978	69	18.826	3.992	−0.558	0.579
Female/Male	8/12			32/37
BDI	20	27.55	10.590
D. SARA
Age	92	29.554	12.760	60	29.85	14.003	−0.660	0.511
Female/Male	69/23			41/19
MADRS	92	26.859	7.117

SD: Standard Deviation; MADRS: Montgomery–Åsberg Depression Rating Scale; BDI: Beck’s Depression Inventory

We next examined if diagnosis impacted brain aging (Table S2). The overall GLM was significant (R²=0.036, F=4.968, df1=4, df2=535, p<0.001) and the MDD group exhibited significantly lower brain centile scores than HC (β=−0.055, SE=0.025, t=−2.194, p=0.029, partial f²=0.007). Post hoc analyses stratified by sex indicated this effect was likely driven by females, although the trend is non-significant (Figure 1A–B). We also carried out post hoc analyses to compare brain centiles of HC versus single episode and recurrent MDD (Table S3). Only the recurrent MDD group showed significantly lower brain centiles than HC (β=−0.058, SE=0.028, t=−2.094, p=0.037, partial f²=0.007; Figure 1C), indicating our initial finding was likely driven by brain centile scores in recurrent MDD. We further tested whether the cumulative exposure to MDD could influence brain aging (Table S4); past depressive episodes (R²=0.087, F=5.571, df1=3, df2=176, p=0.001; β=−0.006, SE=0.005, t=−1.215, p=0.226), current episode duration (R²=0.102, F=6.601, df1=3, df2=174, p<0.001; β=0.001, SE=0.001, t=1.535, p=0.127), and age of MDD onset (R²=0.077, F=4.967, df1=3, df2=178, p=0.002; β=−0.001, SE=0.002, t=−0.324, p=0.746) did not demonstrate a significant relationship with brain centiles. Additionally, while the model was significant (R²=0.030, F=3.100, df1=3, df2=300, p=0.027), brain centile scores did not significantly correlate with depression severity in both single episode and recurrent MDD (Table S5; β=0.015, SE=0.016, t=0.926, p=0.355; Figure 1D), as well as only the recurrent MDD group (β=0.010, SE=0.019, t=0.535, p=0.594). In summary, MDD had atypical brain centile scores relative to HC that were not associated with cumulative MDD exposure and depression severity.

Figure 1:

Both female (A) and male participants (B) diagnosed with major depressive disorder (MDD) did not exhibit significantly different brain centile scores relative to non-depressed controls. Recurrent MDD diagnosis showed significantly lower brain centiles than the control group (C). Depression severity quantified by the Montgomery–Åsberg Depression Rating Scale (MADRS) or Beck’s Depression Inventory (BDI) showed no significant associations with brain centiles (D) Brain centile scores are adjusted for sex and age; the error bars represent the 95% confidence interval. *p < 0.05

Next, we investigated which global brain measures contributed to altered brain centile scores in MDD (Table S6). We performed these analyses separately as sex differences are consistently reported in the brain aging literature(45–47), including in global brain measures driving atypical brain centile scores in neuropsychiatric disorders in our initial report(5). In females (Figure 2A/C/E/G), MDD was significantly associated with a decrease in GMV (β=−13451.3, SE=4981.2, t=−2.70, FDR-p=0.029, partial f²=0.04; Figure S2) and WMV (β=−13091.7, SE=5371.4, t=−2.437, FDR-p=0.046, partial f²=0.02), but not sGMV (β=−898.75, SE=521.09, t=−1.699, FDR-p=0.090) and VV (β=472.81, SE=643.66, t=0.735, FDR-p=0.463). No significant differences were observed by diagnosis in males (Figure 2B/D/F/H).

Figure 2:

Female participants diagnosed with major depressive disorder (MDD) have significantly lower gray matter (A) and white matter volume (C) than non-depressed controls. This was not seen for subcortical (E), and ventricular volumes (G), nor in the male cohort (B, D, F, H). Volumes are adjusted for age; error bars represent the 95% confidence interval. *p < 0.05

We identified which neurocognitive domains were impacted by brain aging in MDD. To constrain our analysis, we first tested the relationship between brain centiles and cognitive performance in HCs (Table S7). After correcting for multiple comparisons, only working memory was significantly associated with brain centile score (β=26.343, SE=7.442, t=3.540, FDR-p=0.005, partial f²=0.09, Figure 3A). Processing speed also showed a similar trend; however, this relationship did not survive correction (β=19.109, SE=7.683, t=2.487, uncorrected-p=0.014; Figure 3B). Based on this, we decided to further analyze these two cognitive domains in all participants (Table S8). Because of normality issues impacting GLM assumptions, we rank transformed our dependent variables. The model for processing speed performance (R²=0.035, F = 2.722, df1=5, df2=379, p=0.020) revealed a significant main effect for brain centile score (β=79.116, SE=29.858, t=2.650, p=0.008, partial f²=0.02), but no significant diagnoses*centile score interaction (β=−49.856, SE=39.992, t=−1.247, p=0.213). Likewise, the GLM predicting working memory performance (R²=0.052, F=3.939, df1=5, df2=357, p=0.002) also had a significant main effect for brain centiles (β=111.383, SE=27.676, t=3.884, p<0.001, partial f²=0.04). In this case, there was a significant two-way interaction, such that there was no significant relationship between working memory (β=−82.790, SE=38.706, t=−2.139, p=0.033, partial f²=0.01; Figure 3C).

Figure 3:

In non-depressed controls, working memory scores (A) showed significantly positive associations with brain centiles while processing speed scores (B) did not. In working memory specifically, significant interaction is seen between brain centiles and diagnosis (C). Neurocognitive domain scores are adjusted for age and sex. *p < 0.05

Subsequently, we wanted to explore whether the presence of CM explained any variance in brain centile scores for individuals with MDD or HC (Table S9). We found no statistically significant prediction of brain centile with overall CM (β=0.009, SE=0.010, t=0.855, FDR-p=0.605; Figure 4A–B), emotional abuse (β=−0.014, SE=0.024, t=−0.586, FDR-p=0.605), physical abuse (β=0.041, SE=0.022, t=1.871, FDR-p=0.248, uncorrected-p=0.031), or sexual abuse (β=0.012, SE=0.023, t=0.571, FDR-p=0.605).

Figure 4:

No significant differences are seen between brain centile scores for varying childhood maltreatment scores, for both the control (A) and major depressive disorder (MDD) group (B). Brain centiles are adjusted for age, sex, diagnosis, and interaction between abuse severity and diagnosis. Also, no significant correlation is seen between escitalopram responses and brain centiles of the major depressive disorder group (C). Percentage changes in Montgomery–Åsberg Depression Rating Scale (MADRS) scores are adjusted for age, sex, and initial depression severity. *p < 0.05

Finally, we assessed whether brain centile scores predict antidepressant response to eight weeks of open-label escitalopram (Table S10). While there was a significant effect of time (β=−13.502, SE=0.574, t=−23.514, p<0.001), there was no significant main effect of centile score (β=0.552, SE=2.033, t=0.271, p=0.786) nor an interaction between centile score and time (β=0.560, SE=2.101, t=0.267, p=0.790). Antidepressant response as MADRS percent improvement (R²=0.023, F=0.945, df1=4, df2=162, p=0.440) did not have a significant main effect (β=1.338, SE=9.344, t=0.143, p=0.886; Figure 4C). A similar trend was observed for response as a dichotomous variable (model R²=0.013, χ²=3.004, df=4, p=0.557; main effect β=0.581, SE=0.591, t=0.984, p=0.325).

Discussion

Brain centile scores provide a quantitative lens through which to analyze the complex interactions between aging and MDD in the hopes of gaining a deeper understanding of the condition’s impact on cognition, treatment response, and the role of risk factors like childhood maltreatment. Here, we showed that people diagnosed with MDD from adolescence to late adulthood exhibited significantly lower brain centile scores relative to non-depressed controls. However, depression severity did not significantly correlate with centile scores within the depressed group. Individuals with recurrent MDD had the largest deviations in brain aging relative to non-depressed controls. Deviations in brain centile score were driven by abnormalities in gray matter and white matter volume, which was most prominent for females and not males. Additionally, we found that centile scores were significantly associated with working memory in non-depressed controls, and this relationship was not present in participants with MDD. CM and improvement in depressive severity after escitalopram treatment were not significantly correlated with brain centile scores. The results offer novel insights into age-dependent deficits in MDD, harnessing aging trajectories from the largest normative dataset to date.

Consistent with our initial hypothesis, we found a lower brain centile score in MDD individuals than in healthy controls, which was associated with lower gray matter and white matter volume in female MDD relative to controls, possibly indicative of accelerated aging(48–50). These results are consistent with previous research using other prediction models that there is a small but significant change in brain age (brain-PAD) in MDD relative to controls(18–20,34). Contrarily, a previous CAN-BIND report revealed no significant differences in baseline brain-PAD between MDD and controls(22). This may be attributed to differences in sample size, normative reference sample, brain-based features, or modeling choices. First, the previous report included solely CAN-BIND-1 participants, while this report used data from two additional datasets, bolstering our sample size and increasing our age range. Second, the original reference sample was 45,615 individuals aged 3–96 years, while Brain Charts uses scans from 95,536 individuals from 115 post-conception to up to 100 years old. Third, our previous report used Freesurfer-generated volume, surface area and cortical thickness values from the Human Connectome Project atlas. The BrainChart method uses volumetric measures for tissue classes and is not stratified by region (except cortical and subcortical gray matter volume). Lastly, although brain-PAD and BrainChart both attempt to assess individual deviation, they have several methodological differences. Brain-PAD compares an individual’s estimated brain age to their chronological age using a machine learning model based on linear gradient tree boosting, and tuned using five-fold cross-validation(51). BrainChart instead uses Generalized Additive Models for Location, Scale and Shape models, which incorporate linear and non-linear trends in volume related to age. Both procedures generated models stratified by sex.

Depression severity in both MDD and more specifically the recurrent group, was not associated with deviations in brain aging, yet some(20), but not all (19,23) previous studies report a positive correlation between brain-PAD and depression severity. Future studies should clarify the relationship between brain aging and MDD severity in light of the notion that MDD is not a unitary disease(52–54); MDD is a clinically heterogeneous disorder. It may be the case that specific symptom profiles are associated with accelerated brain aging.

We also observed sex differences in the relationship between brain centile score inputs and diagnosis. GMV was significantly reduced with an MDD diagnosis, but only in females, providing further evidence of sex differences in brain-based markers of MDD(55). These findings are consistent with previous research regarding sex-specific brain structural changes in MDD. In one study, females exhibited GMV reduction, specifically in the left lingual gyrus and dorsal medial prefrontal gyrus(56).

Brain centile score was positively correlated with working memory in non-depressed controls; greater neurocognitive performance was associated with a higher brain centile score. This finding follows previous literature generally demonstrating working memory impairment with old age(57–59); yet, it is novel in elucidating the relationship between working memory and specifically normative brain aging. Conversely, unlike controls, participants with MDD did not have a significant relationship between working memory performance and brain centile score. It appears that an MDD diagnosis may disrupt age-related effects that normally shape the positive relationship seen in the control group. This result is consistent with our expectation that working memory is disrupted in MDD relative to controls, which has also been well-documented in the literature(60–62). Further studies are needed to clarify if an MDD diagnosis influences the protective effect of a high brain centile score.

We did not find a significant correlation between brain aging and CM; this was inconsistent with our hypothesis and previous studies. In one study, early childhood maltreatment displayed associations with reduced hippocampal GMV(63). Furthermore, a study using the PROCAN data demonstrated that volumes of the amygdala nuclei mediated the severity of depression and anxiety symptoms in at-risk individuals (a cohort that was not included in the present analysis)(64). A recent study also illustrated that sexual abuse during childhood was correlated to a significantly reduced GMV in the right middle occipital gyrus(65). These studies show reductions in GMV as commonly seen in normative brain aging. The lack of expected association may be due to differences in measures for childhood maltreatment between the CAN-BIND studies. On the other hand, emotional subtypes of CM have been suggested as stronger predictors for MDD than physical CM(66–68), which is not found in our analysis. Future studies could also consider other MDD risk factors that are either influenced by CM, such as personality traits, coping styles (69,70), or commonly co-occur with CM, including early-life socioeconomic status(71,72) and parental separation(73,74).

Similarly, our hypothesis regarding escitalopram response was not supported by the findings, which is consistent with the prior CAN-BIND-1 analysis(22). However, other studies revealed that accelerated brain aging was associated with a change in depression severity as measured by the 17-item Hamilton Depression Rating Scale; specifically, a decreased response to 8-week sertraline treatment(23) and an increased response to placebo neuromodulation(20). Therefore, future studies should explore whether brain centiles could predict responses to other antidepressant types of the selective serotonin reuptake inhibitor class or even MDD treatments like transcranial magnetic stimulation. There could also be an exploration into the longitudinal effects of MDD treatment on brain centiles.

We note several limitations of this study. First, our analysis is limited by its cross-sectional nature, which leads to the inability to establish a causal relationship between brain aging and MDD. Prospective data recruited could help to resolve the relationship. For example, one future study could recruit participants who recently experienced a major negative life event like trauma, and determine whether brain age at the time of the event or longitudinal changes in brain age increase the risk for post-traumatic symptoms(75). Second, as previously mentioned, the symptoms of MDD are heterogeneous, and many different symptoms can still lead to a diagnosis. Some recent studies have used functional MRI to identify MDD subtypes based on connectivity profiles in the brain, which could potentially be integrated into brain centiles(76). Third, BrainChart estimates brain age using global brain measures instead of regions of interest. However, volumetric loss in the dorsolateral prefrontal cortex, insula, and hippocampus have been indicated in recurrent MDD(77,78). Exploring centile scores through structures of specific regions of interest is thus an area for future exploration. Additionally, although our sample size is smaller than other studies like the UK Biobank, our MDD group is robust as it contains individuals with a confirmed—and not probable—major depressive episode. Lastly, we have not examined whether brain centiles could reflect longitudinal changes in neurocognitive domain scores with aging or MDD treatment. Longitudinal studies will also help to uncover the directionality of effects; for example, whether atypical brain aging is a cause or consequence of MDD recurrence.

In conclusion, this paper attempted to examine the use of brain centiles as a tool to characterize brain aging and its relationship with MDD diagnosis, cognition, CM, and escitalopram response. We provided evidence substantiating the clinical utility of brain centiles as a predictor of MDD diagnosis and possibly long-term working memory performance. Future studies would need to address general, withstanding issues in the field of brain aging, such as defining causal relationships between brain aging and MDD and incorporating MDD subtypes to consider the heterogeneity of the disorder.