Study Design

The NeuroCure (DFG EXC 257)-funded study used a retrospective, cross-sectional design with standardized visits at the Neuroscience Clinical Research Center and the Berlin Center for Advanced Neuroimaging at Charité—Universitätsmedizin Berlin. ELS severity scores and age in years were used as continuous predictors of biomarker, brain, and cognitive outcomes. The study was approved by the institutional ethics committee and was conducted in accordance with the Ethical Principles for Medical Research as established by the Medical Association Declaration of Helsinki. Written informed consent was obtained from all participants.

Sample

We recruited 184 healthy women aged 30 to 60 years to capture an age range during which alterations associated with brain aging occur. We recruited exclusively women from Berlin via advertisements online, in the public transportation system, and across the city. We did not target specific neighborhoods, events, or clinical populations. As explained above, we used a high-risk population approach and focused our study on women. In brief, the rationale for this focus is based on higher AD risk,³⁷ greater vulnerability to stress,³⁸ and a higher risk to develop certain adverse outcomes of ELS⁴¹ in women as compared with men. This targeted approach allows us to yield evidence of the impact of ELS on brain aging within this high-risk group. We acknowledge that this design does not allow us to detect sex-specific vulnerability. Yet, our approach follows established guidelines advocating for targeted research in high-risk populations and in women's health to establish foundational evidence for future expanded studies. Indeed, women may exhibit a specific vulnerability to the effects of ELS on AD risk; however, our study was not designed to test this hypothesis.

Women were excluded if they had a formal physician-derived diagnosis of neurological disorders, including manifested AD and other dementias, as ascertained by neurological examination and anamnesis and by monitoring anti-dementia medication intake. Of note, we did not exclude mild cognitive impairment (MCI), given that early cognitive impairment is one of the primary outcomes of the study. Further exclusion criteria included recent use of immunosuppressants, any medicines with central nervous system (CNS) effects, current pregnancy, and magnetic resonance imaging (MRI) contraindications. Current psychiatric disorders were exclusionary, with the exception of depression, anxiety disorders, and post-traumatic stress disorder (PTSD) due to their high prevalence in individuals with ELS and the potential for a mediating relationship between ELS and outcomes. Excluding affective disorders related to ELS would have resulted in a highly selective sample characterized by specific resilience to ELS, which may also affect resilience to dementia development. Our final sample consisted of 179 women who participated in the study between March 2021 and September 2022. All participants reported their sex and gender as female. Sociodemographic and clinical characteristics of the sample are presented in Table 1.

Assessment of Early-Life Stress

We included women with and without ELS exposure. Presence or absence and severity of ELS were assessed using the German version of the Comprehensive Trauma Interview (CTI),⁴³ a psychometrically validated tool demonstrating good inter-rater reliability with kappa values ranging from 0.51 to 0.76⁴⁴ and good predictive validity for psychiatric outcomes.⁴³ In our study, the CTI demonstrated a high level of internal consistency with a Cronbach's alpha of 0.84. For validation purposes, we used the Childhood Trauma Questionnaire (CTQ)⁴⁵ and the correlation between CTI scores and CTQ scores in our sample was 0.77 (p < 0.001), demonstrating good convergent validity.

Exposure levels, that is, severity scores, were determined by counting items corresponding to the US Centers for Disease Control and Prevention definition of adverse childhood experiences (ACE)⁴⁶ in the CTI and included emotional, sexual, and physical abuse, emotional and physical neglect, witnessing domestic violence, household substance abuse, family psychiatric disorders or suicide attempt, or death by suicide, imprisonment of family member, instability due to parental separation/divorce, and death of a primary caregiver. The ACE score, ranging from 0 to 11, was calculated for each participant as the sum of the number of experienced stressors before menarche (see Table 1). For detailed score distributions, see Supplementary Figures S1 to S3.

We used menarche as a developmental cutoff to define the early life period, as opposed to a certain cutoff age. Our approach to define early life by a developmental stage rather than by a specific age is rooted in the available literature regarding sensitive periods for stress effects.^{47, 48} Notably, hormonal changes occurring with the onset of puberty impact both stress reactivity and neuroplastic responses.^{49, 50} Stress before and after menarche has differential long-term effects on brain structure and function.⁵¹ Puberty marks a transition where the focus of the child shifts from needs expected from caretakers to needs related to peers.⁵² Importantly, using menarche as a developmental cutoff to define the early life period accounts for individual variability in pubertal timing with impact on brain outcomes. Of note, it has been reported that ELS may impact pubertal development. To exclude potential effects of ELS exposure on age at menarche, we assessed associations between age at menarche, ELS scores, and all outcome variables. The mean age of menarche in our sample was 12.98 years (standard deviation [SD] = 1.27). No significant correlations were found between age at menarche and ELS scores or any outcome variables (see Supplementary Materials for details).

Blood Sampling and Assays

We assessed concentrations of blood-based biomarkers of neuroinflammation and neurodegeneration using both serum and plasma. Serum measures were used for glial fibrillary acidic protein (sGFAP), neurofilament light chain (sNfL) and tau phosphorylated at threonine 181 (sp-tau181). Amyloid-β 42 (pAβ42) concentrations were assessed in plasma.

These blood-based biomarkers reliably reflect protein levels in the CNS: GFAP is an astrocytic protein expressed in the CNS. Elevated peripheral levels of GFAP in plasma or serum are indicative of astroglial activation or injury within the CNS. Thus, peripherally measured GFAP serves as a marker of neuroinflammation.⁵³ NfL is a structural protein expressed predominantly in neurons. Its presence in peripheral blood correlates with axonal damage or degeneration, confirming that it is a reliable biomarker for neurodegenerative processes and CNS injury.⁵⁴ P-tau181 is a marker associated with neurodegenerative diseases, particularly AD and other tauopathies. Changes in p-tau181 levels in blood reflect pathological processes occurring in the CNS, including neuroinflammation and tau-related neurodegeneration.⁵⁵ Aβ42 is a peptide involved in the formation of amyloid plaques, a hallmark of AD. Blood levels of Aβ42 are validated to reflect amyloid pathology within the CNS.⁵⁶ The use of serum for GFAP, NfL, and p-tau181 as opposed to plasma is supported by studies demonstrating strong correlations between serum and plasma levels of these biomarkers.^{57, 58} Specifically p-tau181 in serum has been extensively validated as comparable to plasma p-tau181 across different assay types, including the Quanterix Single-Molecule Analysis Technology (SIMOA; Simoa) assay used in this study, with robust separation between diagnostic groups and strong correlations with CSF measures.^{59, 60}

Venous blood was collected in EDTA and serum tubes (BD Vacutainer) and centrifuged. Serum samples were allowed to clot for 30 minutes at room temperature before centrifugation. Plasma and serum samples were aliquoted and stored at −80°C. The sGFAP (n = 10 missing), sNfL (n = 8 missing), and sp-tau181 (n = 9 missing) levels were measured using SIMOA at Labor Berlin—Charité Vivantes GmbH, using commercially available kits (Quanterix, Billerica, MA, USA). For the SIMOA assay, calibrators and high versus low controls were run in duplicates as quality controls, whereas samples were run in singlicates. The intra-assay coefficients of variation (CV) were 2.05% for sGFAP and 7.45% for sp-tau181, whereas singlicate measures are formally accredited by Labor Berlin for sNfL. Missing data pertain to blood sampling failure. The pAβ42 was measured using a commercial Human Aβ42 Enzyme Linked Immunosorbent Assay (ELISA) kit (Invitrogen, Thermo Fisher Scientific, Henningsdorf, Germany). The intra-assay CV was 20.2% and the inter-assay CV was 14.9%.

MRI Acquisition and Data Preprocessing

Structural MRI was performed using a 3 T PRISMAFit MRI scanner (Siemens Healthineers, Erlangen, Germany) with a standard 64-channel head coil. Three-dimensional T1-weighted (T1w) magnetization prepared rapid acquisition of gradient echo (MPRAGE) and T2-weighted (T2w) SPACE images were acquired for all participants using the following scan parameters: T1w TE = 2.22 ms, TR = 2,400 ms, flip angle = 8, bandwidth = 220 Hz/Px, FOV = 256 mm, resolution 0.8 mm isotrope, and TA = 6:38 minutes (T2w) TE = 563 ms, TR = 3,200 ms, bandwidth = 744 Hz/Px, FOV = 256 mm, resolution 0.8 mm isotrope, TA = 5:57 minutes. Images of 172 participants were organized according to the Brain Imaging Data Structure (BIDS) standard. For 7 women, imaging protocols were incomplete or cancelled mid-scan. The T1w images were corrected for intensity non-uniformity with N4BiasFieldCorrection,⁶¹ distributed with ANTs 2.3.3⁶² (RRID: SCR_004757), and used as T1w-reference throughout the workflow. The T1w-reference was then skull-stripped with a Nipype implementation of the antsBrainExtraction.sh workflow, using OASIS30ANTs as the target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM), and gray matter (GM) was performed on the brain-extracted T1w using fast (FSL 6.0.5.1:57b01774 and RRID:SCR_002823). Brain surfaces were reconstructed using recon-all applied to the T1w and the T2w images (FreeSurfer 6.0.1 and RRID:SCR_001847⁶³). The preprocessing pipeline was run using functional magnetic resonance imaging (fMRI)prep focusing on the structural part within the individual subject space.⁶⁴ We extracted volume data of the following global structures: total brain volume (TBV; all segmented brain tissue without ventricles, CSF, and choroid plexus), total gray matter volume (tGMV), total white matter volume (tWMV), subcortical gray matter volume (sGMV), and lateral, third, and fourth ventricular volumes, which serve as indicators for dementia risk (see above) as well as intracranial volume (ICV).

Assessment of Cognitive Function

Cognitive function was assessed using tasks from the Cambridge Neuropsychological Test Automated Battery⁶⁵ (CANTAB). Tasks sensitive to the development of cognitive impairment that may be observed in early stages of AD, including MCI, were selected. Tasks included the Spatial Working Memory task (SWMS), which tests executive function and measure of strategy as well as working memory errors. This task is impacted by visuo-spatial information retention-deficits often observed in early stage AD. The Rapid Visual Information Processing (RVPA) task, which provides a measure of sustained attention, and the pattern recognition memory percent correct instant (PRMPCI) task, which tests visual pattern recognition memory in a 2-choice forced discrimination paradigm, has been shown to be particularly sensitive to temporal lobe dysfunction. Tasks were presented in fixed order to all participants as follows: (1) PRMPCI (n = 1 missing), (2) SWMS (n = 6 missing), and (3) RVPA (n = 5 missing). For details on tasks, see the Supplementary Material. Of note, we did not use the CANTAB for diagnostic purposes, and formal diagnoses of prodromal AD or MCI were not made based on task performance. The CANTAB does not provide standardized clinical cutoffs for diagnostic classification.

Psychiatric and Socioeconomic Assessment

Psychiatric diagnoses were assessed by standardized clinical interview.⁶⁶ Current depressive symptoms were assessed using the German version of the Beck's Depression Inventory II (BDI-II).⁶⁷ Current symptoms of PTSD were assessed using the German version of the short screening scale for DSM-IV.⁶⁸ Anxiety was assessed using the German version of the Spielberger State–Trait Anxiety Inventory (STAI).⁶⁹ Socioeconomic status (SES) was defined as a combination of educational level (assessed in categories from less than high school to advanced degrees and recoded into values from 1 through 5) and household monthly income (assessed in categories from ≤ 100€ to ≥ 20,000€ and recoded into values from 1 through 5).⁷⁰ These measures were used as covariates.

Statistical Analysis

Biomarker concentrations were natural-log-transformed for analysis, as concentrations were not normally distributed. Global GM and WM and ventricular volumes were adjusted by ICV. Bivariate Pearson correlations were computed to inspect linear associations between all dependent variables (DVs), age, and ACE score. Linear models, with an interaction term between age and ACE score, were implemented to assess heteroscedasticity and normality (linear model). These models are provided in the Supplementary Table S1. Next, separate GAMs were implemented for all DVs to investigate the nonlinear interaction effects between ACE score and age in the prediction of outcomes (model 1). Tensor product smooth terms were utilized to model this nonlinear interaction effect with default smoothing functions (thin plate regression splines which are the default in the mgcv package in R), and model fitting was performed using the Restricted Maximum Likelihood method. Goodness of fit for model 1 was tested in comparison to a GAM with a tensor product smooth term, capturing the interaction effect stratified by the levels of age for all DVs (model 2) as several DVs correlated linearly with age. Model fit was assessed using a combination of Akaike Information Criteria, Bayes Information Criteria, residual deviance, and residual degrees of freedom (see Supplementary Table S2). The model that best balanced the theoretical expectations and relevant fit criteria was retained as the main model, that is, model 1. Significant smooth terms for the nonlinear interaction effects between age and ACE score in predicting outcomes in each of the models indicate that the relationship between these predictor variables and DVs is not adequately captured by a simple linear relationship.

As pAβ42 concentrations were below the ELISA detection limit in 79 women, the participants were stratified into 2 groups by ELS exposure, that is, ELS− (≤ 2 ACE) versus ELS+ (≥ 3 ACE), and an exploratory analysis was conducted using a Chi-square test to assess for between-group differences in detectability.

We assessed the impact of symptom severity of depression, state and trait anxiety, PTSD, and SES by integrating these variables as linear covariates into the main model. Bivariate Pearson correlations were conducted among these covariates, age, and ACE scores (Supplementary Table S3). The model performance was compared with and without covariates (Supplementary Table S4). Variance inflation factor (VIF) analyses were conducted to evaluate multicollinearity.

All statistical analyses were run in R Project for Statistical Computing (R Core Team, Vienna, Austria; version 4.1.2, 2021) and IBM SPSS Statistics (IBM, version 29, 2022). GAMs were implemented using the “mgcv” package for R.⁷¹ All analyses account for multiple comparisons conducted within each domain and p value adjustments were applied using the Benjamini-Hochberg (false discovery rate [FDR]) method. All reported p values in the Results section represent the adjusted p values. The significance level was set at p < 0.05 for all analyses.

Descriptives

Bivariate correlations of all predictors and DVs are depicted in Table 2. Age was positively associated with sNfL concentrations, sGFAP concentrations, lateral and third ventricular volumes, and CANTAB SWMS. Age was negatively associated with TBV, tGMV, sGMV, tWMV, and CANTAB RVPA and PRMPCI. Age was not associated with sp-tau181 levels or fourth ventricular volume. There were no linear correlations between ACE score and any of the DVs. For the subsequent nonlinear GAM analyses, model 1 was used for all DVs, as described above.

Associations of ELS Severity and Age With Blood-Based Biomarker Concentrations

Levels of all peripheral biomarker concentrations correlated significantly with one another (see Table 2). GAM analysis revealed a significant non-linear interaction effect between age and ACE score in predicting sGFAP and sNfL concentrations (Table 3). The model for sGFAP indicates a more pronounced age-dependent increase in sGFAP concentrations with higher ACE exposure levels (Fig 1A). The model for sNfL shows a U-shaped curve, with highest predicted sNfL concentrations at high (7+) ACE exposure levels and lowest predicted sNfL concentrations at low to moderate (1–3) ACE exposure levels (Fig 1B). There was no statistically significant interaction effect between age and ACE score in predicting sp-tau181 (see Table 3). The exploratory analysis on detectability of Aβ42 in plasma revealed that pAβ42 was more frequently detectable in the ELS+ (55%) than in the ELS− group (45%; χ²(2) = 5.57, p = .018).

Associations of ELS Severity and Age With Global Brain and Ventricular Volumes

GAM analysis revealed that the smooth terms for the nonlinear interaction effect between age and ACE score in predicting TBV, tGMV, sGMV, and third ventricular volume were significant (see Table 3). The models for both TBV and tGMV show a more pronounced decrease in volume with increasing age and higher ACE exposure level (Fig 2A, B). Interestingly, the model for sGMV shows an increase in volume with increasing ACE exposure level at a younger age and a decrease, that is, faster deterioration, in volume with increasing ACE exposure level at an older age (Fig 2C). The model for the third ventricular volume predicts the largest ventricle size at the highest age and highest ACE exposure level (Fig 2D). There were no significant interaction effects between age and ACE score in predicting tWMV or lateral or fourth ventricular volumes.

Associations of ELS Severity and Age With Cognitive Function

GAM analysis revealed a significant interaction effect between ACE score and age in predicting RVPA, SWMS, and PRMPCI (see Table 3). The model for SWMS predicted the highest scores at low ACE exposure level and young age and the lowest scores at higher ACE exposure level and older age (Fig 3A). The model for RVPA indicates a more pronounced cognitive dysfunction with higher ACE exposure level at a younger age (Fig 3B). The model for PRMPCI indicates a more pronounced age-dependent cognitive impairment with higher ACE exposure level (Fig 3C).

Adjustment for Psychiatric Symptoms and Socioeconomic Status

Bivariate correlation analysis revealed significant correlations among all 4 psychiatric symptom measures, SES, and ACE scores (see Supplementary Table S3). VIF analysis indicated multicollinearity of these variables with ACE scores. When including these covariates (psychiatric symptoms and SES) in the main model (as calculated in model 3), we found that the above effects remained statistically significant for 13 of 15 outcomes (see Supplementary Table S4). For 2 cognitive outcomes (RVPA and PRMPCI), effects were attenuated by SES, but remained at a statistical trend level. Psychiatric symptoms as covariates had no effect in model 3.