Transient Elevation and Subsequent Normalization of Brain Age Gap During the Postpartum Period Authors: Chang‐hyun Park, Yunjin Bak, Sanghoon Han, Seung‐Koo Lee, Na‐Young Shin Published: 2026 ABSTRACT ABSTRACTThe postpartum period involves substantial brain changes, but whether these represent pathological aging or adaptive plasticity remains unclear. We applied machine learning‐based brain age prediction to quantify postpartum brain aging trajectories in a controlled longitudinal study. Brain age gap, the difference between MRI‐predicted and chronological age, was assessed in 55 women (25 postpartum and 30 nulliparous controls). Postpartum women were assessed at approximately 3 and 12 months postpartum; controls were assessed at matched intervals. Postpartum women showed elevated brain age gap at 3 months (2.5 ± 2.7 years) compared with controls (0.3 ± 3.6 years; p = 0.01, Cohen's d = 0.69), which normalized by 12 months. Within‐group analysis revealed significant brain age gap decrease in postpartum women (−3.3 ± 3.5 years; p = 0.008), while controls showed minimal change; the between‐group difference in longitudinal change did not reach statistical significance (p = 0.12, Cohen's d = −0.58). Across all participants, brain age gap correlated with working memory performance (ρ = −0.30; p = 0.03), depressive symptoms (ρ = 0.38; p = 0.004), and subjective cognitive complaints (ρ = 0.28; p = 0.04) at baseline, providing preliminary evidence of clinical relevance. Explainability analysis using Shapley additive explanations identified subcortical structures as the network most selectively associated with the postpartum elevation in brain age gap at 3 months (p = 0.006, Cohen's d = 0.82), with normalization by 12 months. Together, these findings indicate transient brain age gap elevation in postpartum women without statistical confirmation of a trajectory differing from controls. With brain age gap providing a quantitative index of postpartum brain structural change, these preliminary findings are consistent with adaptive neuroplasticity, suggesting that such change may be temporary rather than reflecting permanent impairment.Keywords: brain age gap, brain age prediction, maternal brain, postpartum, Shapley additive explanations Introduction The transition to motherhood represents a significant neurobiological transformation involving extensive hormonal fluctuations and behavioral adaptations that fundamentally reshape brain structure and function. Neuroimaging studies have widely documented substantial brain structural changes during pregnancy and postpartum, most notably gray matter reductions and white matter alterations (Hoekzema et al. 2017; Paternina‐Die et al. 2024). These alterations were initially interpreted as potential vulnerability markers. Yet, accumulating evidence demonstrates the adaptive and reversible nature of maternal brain changes (Hoekzema et al. 2017; Kim et al. 2010; Pritschet et al. 2024; Paternina‐Die et al. 2024). Such reversibility is not unique to pregnancy; comparable patterns of transient structural change have been reported across other female endocrine transitions, including the menstrual cycle (Lisofsky et al. 2015) and menopause (Mosconi et al. 2021). Nevertheless, the temporal dynamics of postpartum structural recovery remain incompletely characterized. While some brain changes show progressive recovery over the postpartum months (Kim et al. 2010; Martínez‐García, Paternina‐Die, Desco, et al. 2021), others have been reported to persist for years after childbirth (Hoekzema et al. 2017; Martínez‐García, Paternina‐Die, Barba‐Müller, et al. 2021), raising questions about the timeline and completeness of postpartum brain normalization. Longitudinal investigations have revealed that brain structural changes during the postpartum period follow distinctive nonlinear trajectories, including U‐shaped patterns characterized by initial gray matter decreases followed by progressive recovery (Hoekzema et al. 2017; Luders et al. 2020), with more recent work linking these patterns to hormonal changes and maternal attachment (Servin‐Barthet et al. 2025). However, these regional structural changes are complex and distributed across multiple brain systems, making it challenging to quantify their overall impact on brain health and to determine whether they represent elevated aging or adaptive reorganization. Machine learning‐based brain age prediction offers a powerful approach to address this challenge by quantifying brain aging status and deviations from normative trajectories (Cole and Franke 2017; Franke and Gaser 2019). This approach estimates biological brain age from structural MRI features, with brain age gap (BAG), defined as predicted age minus chronological age, serving as a brain health index. Accumulating evidence demonstrates BAG's clinical utility, revealing robust associations with mortality risk in older populations (Cole et al. 2018) and with neuropsychiatric conditions and cognitive outcomes across a broad age range (Kaufmann et al. 2019). In initial findings from applying brain age prediction to the maternal brain, Luders et al. (2018) reported BAG decrease during early postpartum weeks, suggesting rapid structural recovery. Despite this promising initial work, important questions persist. First, the absence of a control group precluded determining whether observed BAG changes represented recovery from an initially elevated state. Second, the functional significance of these BAG changes was not examined, leaving unresolved whether BAG captures functionally relevant neurobiological variations during the postpartum period. Here, we characterize postpartum brain aging trajectory using BAG in a controlled longitudinal design. We first examine longitudinal BAG changes from 3 to 12 months postpartum and which neural systems contribute most to observed BAG differences using explainability analysis. Additionally, we investigate preliminary associations of BAG with working memory, depressive symptoms, and subjective cognitive complaints across postpartum women and nulliparous controls. We aimed to test the hypothesis that postpartum women show a biphasic trajectory with initially elevated BAG relative to nulliparous controls, followed by progressive normalization. Methods Study Design and Participants We conducted a prospective longitudinal study at a single academic medical center. Eligibility criteria for all participants included age 20–40 years, right‐handedness, and absence of neurological or psychiatric disorders meeting Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition criteria. Postpartum eligibility additionally required delivery at 37 weeks' gestation or later, uncomplicated delivery, and absence of severe obstetric complications. Controls were required to have no prior history of pregnancy. Exclusion criteria included psychotropic medication use, head injury with loss of consciousness exceeding 5 min, MRI contraindications, and substance abuse. The study included 25 postpartum women (mean age 30.9 ± 3.0 years) and 30 nulliparous controls (mean age 29.1 ± 4.2 years). Postpartum women were assessed at approximately 3 months (baseline, mean time postpartum 3.4 ± 0.5 months) and 12 months (follow‐up, mean time postpartum 12.9 ± 0.7 months) postpartum (mean follow‐up interval 9.5 ± 0.6 months). Nulliparous controls underwent baseline and follow‐up assessments at matched intervals (mean follow‐up interval 10.4 ± 2.9 months). Of 25 postpartum women, 11 (44%) completed follow‐up; of 30 controls, 20 (67%) completed follow‐up. Loss to follow‐up was primarily due to subsequent pregnancy in postpartum women and marriage or pregnancy in controls. The institutional review board approved the protocol, and all participants provided written informed consent. Clinical and Cognitive Assessment At baseline, participants completed comprehensive cognitive testing, self‐report measures, and standardized mood assessments. Working memory was assessed using the digit span sequencing subtest from the Wechsler Adult Intelligence Scale‐Fourth Edition (Wechsler 2008). Mood was assessed with the Beck Depression Inventory‐Second Edition (BDI‐II; Beck et al. 1996). Subjective cognitive complaints were measured with a validated 45‐item questionnaire covering memory, attention, and executive function domains (McNair and Kahn 1983). Additional cognitive measures included verbal fluency (animal naming) and trail making tests (Parts A and B; Reitan 1958) to assess semantic fluency, processing speed, and executive function. Blood samples were collected at baseline in the morning following overnight fasting for serum estradiol measurement via electrochemiluminescence immunoassay. Genomic DNA was extracted from peripheral blood for apolipoprotein E (APOE) genotyping to determine E2/E3/E4 allele status, with participants classified as APOE E4 carriers based on the presence of at least one E4 allele. MRI Data Acquisition and Preprocessing High‐resolution structural brain MRI data were acquired on a 3‐Tesla MRI scanner (GE Healthcare Discovery MR750, Milwaukee, WI, USA) using a 3D T1‐weighted Spoiled Gradient Recalled Echo sequence with the following parameters: repetition time 8.3 ms, echo time 3.3 ms, flip angle 12°, in‐plane resolution 0.77 × 0.86 mm, slice thickness 1.0 mm, and 216 slices. Structural images were preprocessed using the Computational Anatomy Toolbox (CAT12.9; http://www.neuro.uni‐jena.de/cat/) implemented in Statistical Parametric Mapping (SPM12; https://www.fil.ion.ucl.ac.uk/spm/). The preprocessing pipeline comprised intensity non‐uniformity correction, skull stripping, tissue segmentation into gray matter, white matter, and cerebrospinal fluid, spatial normalization to Montreal Neurological Institute standard space using nonlinear registration, and resampling to isotropic 1.0 mm3 voxel size with dimensions cropped to 157 × 189 × 156 to minimize extra‐cerebral margins while preserving all brain structures. All processed images underwent visual quality control inspection. Brain Age Prediction Model Development We developed a deep convolutional neural network for brain age prediction based on a previously validated architecture (Park et al. 2025). The model processes gray matter, white matter, and cerebrospinal fluid probability maps through a three‐stage architecture: (1) a channel‐wise attention module for tissue‐specific recalibration, (2) a convolutional feature extraction module comprising four 3D convolutional blocks for spatial feature extraction, and (3) a regression module for brain age prediction (Figure S1). The model training workflow comprised three phases. First, the model was pretrained on 639 healthy women from the Human Connectome Project Young Adult and Aging cohorts (Van Essen et al. 2013; Bookheimer et al. 2019), comprising 606 women aged 22–36 years from the Young Adult cohort and 33 women aged 36–40 years from the Aging cohort. Transfer learning was then applied by freezing all pretrained parameters in the attention module and convolutional feature extraction module, while retraining only the regression module on the study sample to prevent overfitting. Second, stratified five‐fold cross‐validation was conducted at the participant level for robust performance evaluation, ensuring all scans from the same individual were assigned to the same fold. Third, a final model was trained on the complete dataset using the same transfer learning procedure with hyperparameters validated through cross‐validation to generate brain age predictions for all scans at both baseline and follow‐up timepoints. Detailed model architecture and training procedures are provided in Supporting Information. Predicted brain age values were bias‐corrected to address systematic age‐dependent prediction errors using established methods (de Lange and Cole 2020). BAG was computed as bias‐corrected predicted brain age minus chronological age, with positive values indicating older‐appearing brain structure and negative values indicating younger‐appearing brain structure. Detailed bias correction procedures are provided in Supporting Information. Explainability Analysis To identify which brain regions contributed most to brain age predictions, we applied Shapley additive explanations (SHAP) (Lundberg and Lee 2017), an explainability framework grounded in cooperative game theory that quantifies each input feature's contribution to model predictions. SHAP values represent the contribution of each brain voxel in units of years to the difference between an individual's predicted brain age and the average predicted brain age across the study population. SHAP values were computed separately for gray matter, white matter, and cerebrospinal fluid tissue probability maps using the entire study sample as the reference distribution. For visualization purposes, mean absolute SHAP values were computed for each group and timepoint to illustrate overall spatial patterns of regional importance. For statistical analysis, voxel‐wise SHAP values from gray matter were aggregated within eight spatially non‐overlapping functional brain networks based on established resting‐state functional connectivity parcellation (Yeo et al. 2011): visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode networks, plus a subcortical network comprising hippocampus, amygdala, basal ganglia, and thalamus. Network‐level SHAP values were computed as the mean of all gray matter voxel SHAP values within each network mask. Signed SHAP values were used to test directional group differences, assessing whether networks contributed toward relatively older versus younger brain age predictions. Detailed SHAP computation procedures are provided in Supporting Information. Statistical Analysis All statistical analyses were conducted using nonparametric methods given non‐normal distributions and modest sample size. Cross‐sectional group comparisons employed Mann–Whitney U tests, within‐group longitudinal changes employed Wilcoxon signed‐rank tests, and clinical associations employed Spearman rank correlations. Effect sizes were quantified using Cohen's d, with thresholds of 0.2, 0.5, and 0.8 interpreted as small, medium, and large effects, respectively. Multiple comparison corrections were applied using the Benjamini–Hochberg false discovery rate (FDR) procedure for three families of tests: longitudinal BAG trajectory comparisons comprising the group comparison at baseline, the group comparison at follow‐up, and the between‐group difference in longitudinal BAG change; functional network SHAP comparisons comprising eight network‐level tests at each timepoint, with FDR correction applied separately within each timepoint; and clinical correlations of BAG with the three primary measures in the overall sample. Within‐group longitudinal BAG changes were reported without additional FDR correction as confirmatory analyses independent of the between‐group trajectory comparisons. Secondary clinical measures and genetic markers were analyzed without FDR correction as exploratory analyses. For all tests, two‐sided p < 0.05 indicated statistical significance. To formally test whether longitudinal BAG trajectories differed between groups, we additionally fitted a linear mixed‐effects model with Group (postpartum vs. control), Time (baseline vs. follow‐up), and their interaction as fixed effects, with random intercepts by participant, restricted to participants with complete longitudinal data. Given that primary analyses employed nonparametric methods, the mixed‐effects model is reported as a sensitivity analysis. Results Sample Characteristics Among 55 participants, groups were well matched for demographic characteristics, with no significant differences in age (control: 29.1 ± 4.2 vs. postpartum: 30.9 ± 3.0 years; Mann–Whitney U test: z = −1.90, p = 0.06) or education (control: 17.8 ± 2.1 vs. postpartum: 17.0 ± 2.5 years; z = 1.39, p = 0.16; Table 1). At baseline, postpartum women showed lower working memory performance assessed by digit span sequencing (control: 9.2 ± 2.1 vs. postpartum: 8.0 ± 2.5; z = 2.11, p = 0.04), higher subjective cognitive complaints (control: 28.1 ± 8.9 vs. postpartum: 34.9 ± 12.9; z = −1.97, p = 0.049), and a trend toward elevated depressive symptoms on BDI‐II (control: 6.8 ± 5.5 vs. postpartum: 9.0 ± 5.9; z = −1.37, p = 0.17). As expected, estradiol levels were significantly lower in postpartum women (control: 134.5 ± 97.7 vs. postpartum: 62.9 ± 43.1 pg/mL; z = 2.91, p = 0.004). APOE E4 carrier prevalence was similar between groups (control: 17% vs. postpartum: 12%; p = 0.72). TABLE 1Baseline participant characteristics.VariableControl (n = 30)Postpartum (n = 25) z p DemographicsAge, years29.1 ± 4.230.9 ± 3.0−1.900.06Education, years a 17.8 ± 2.117.0 ± 2.51.390.16Study timelineTime to follow‐up, months10.4 ± 2.99.5 ± 0.6−0.090.92Postpartum timing at baseline, monthsNA3.4 ± 0.5NANAPostpartum timing at follow‐up, monthsNA12.9 ± 0.7NANAGenetic characteristicsAPOE E4 carrier, n (%) b 5 (17)3 (12)NA0.72Primary measuresDigit span sequencing c 9.2 ± 2.18.0 ± 2.52.110.04BDI‐II d 6.8 ± 5.59.0 ± 5.9−1.370.17Subjective cognitive complaints28.1 ± 8.934.9 ± 12.9−1.970.049Secondary measuresEstradiol, pg/mL e 134.5 ± 97.762.9 ± 43.12.910.004Verbal fluency f 21.0 ± 3.920.5 ± 5.80.700.48Trail making A, seconds f 22.9 ± 7.822.6 ± 8.40.340.73Trail making B, seconds f 61.5 ± 26.654.1 ± 22.41.030.30 Note: Data are presented as mean ± SD unless otherwise indicated. p‐values are from Mann–Whitney U tests for continuous variables and Fisher exact tests for categorical variables.Abbreviations: APOE, apolipoprotein E; BDI‐II, Beck Depression Inventory‐Second Edition; NA, not applicable. a Control: n = 26. b Fisher exact test was used. c Control: n = 29. d Postpartum: n = 24. e Control: n = 28; postpartum: n = 21. f Control: n = 28; postpartum: n = 24. Baseline participant characteristics. Note: Data are presented as mean ± SD unless otherwise indicated. p‐values are from Mann–Whitney U tests for continuous variables and Fisher exact tests for categorical variables. Abbreviations: APOE, apolipoprotein E; BDI‐II, Beck Depression Inventory‐Second Edition; NA, not applicable. Control: n = 26. Fisher exact test was used. Control: n = 29. Postpartum: n = 24. Control: n = 28; postpartum: n = 21. Control: n = 28; postpartum: n = 24. Brain Age Prediction Model Performance The pretrained model achieved mean absolute error of 2.65 years on the Human Connectome Project test set. Following transfer learning adaptation to the study dataset, participant‐level five‐fold cross‐validation achieved mean absolute error of 2.80 ± 0.56 years across validation folds. The final model trained on all 86 scans from 55 participants achieved mean absolute error of 2.68 years. These performance metrics are comparable to other neuroimaging‐based brain age prediction models, indicating adequate model calibration for assessing brain aging patterns in postpartum and control populations. Scatterplots of chronological versus predicted brain age by group are provided in Figure S2. Brain Age Gap by Group and Timepoint At baseline, postpartum women demonstrated significantly elevated BAG (2.5 ± 2.7 years) compared with controls (0.3 ± 3.6 years), a mean difference of 2.2 years (95% confidence interval (CI): 0.6 to 3.8 years; Mann–Whitney U test: z = −2.56, p = 0.01, Cohen's d = 0.69) (Figure 1A, Table 2). By follow‐up, this difference resolved, with postpartum women (0.1 ± 2.9 years) showing BAG values virtually identical to controls (0.1 ± 3.7 years), a mean difference of less than 0.1 years (95% CI: −2.3 to 2.3 years; z = −0.64, p = 0.52, d < 0.01) (Figure 1B). The BAG distributions were consistent with this pattern, showing a rightward shift in the postpartum group at baseline that attenuated by follow‐up (Figure 2). Brain age gap (BAG) trajectory across groups and timepoints. (A) Baseline: Postpartum women showed greater BAG (2.5 ± 2.7 years) compared with controls (0.3 ± 3.6 years; Mann–Whitney U test: z = −2.56, p = 0.01, Cohen's d = 0.69). (B) Follow‐up: Group differences resolved (postpartum: 0.1 ± 2.9 years; control: 0.1 ± 3.7 years; z = −0.64, p = 0.52, d < 0.01). Bars represent group means. Individual data points are shown as jittered circles. Error bars indicate standard deviation. (C) Longitudinal BAG trajectories in participants with complete data (n = 31; 11 postpartum and 20 nulliparous women): Postpartum women showed significant BAG decrease (−3.3 ± 3.5 years; Wilcoxon signed‐rank test: z = −2.67, p = 0.008, d = −0.94); controls showed minimal change (−0.7 ± 4.9 years; z = −0.90, p = 0.37, d = −0.14). Between‐group difference in longitudinal change was non‐significant (Mann–Whitney U test: z = 1.55, p = 0.12, d = −0.58). Individual trajectories are shown as connected circles, with group means shown as thicker lines. *p < 0.05. TABLE 2Brain age gap (BAG) by group and timepoint.AnalysisValue, yearsMean difference, years (95% CI) z p Cohen's d Cross‐sectional group comparisonsBaselinePostpartum2.5 ± 2.72.2 (0.6, 3.8)−2.560.010.69Control0.3 ± 3.6Follow‐upPostpartum0.1 ± 2.9< 0.1 (−2.3, 2.3)−0.640.52< 0.01Control0.1 ± 3.7Within‐group longitudinal comparisonsPostpartum (follow‐up—baseline)−3.3 ± 3.5−3.3 (−5.4, −1.5)−2.670.008−0.94Control (follow‐up—baseline)−0.7 ± 4.9−0.7 (−2.8, 1.5)−0.900.37−0.14Between‐group longitudinal comparisonDifference in change−2.6 (−5.6, 0.3)1.550.12−0.58 Note: Values represent mean ± SD of BAG at each timepoint for cross‐sectional comparisons and mean ± SD of BAG change for longitudinal comparisons. The mean difference column indicates the between‐group difference in BAG for cross‐sectional comparisons, mean BAG change for within‐group comparisons, and the between‐group difference in BAG change for the longitudinal between‐group comparison. Cross‐sectional group comparisons and the between‐group difference in longitudinal change used Mann–Whitney U tests; within‐group longitudinal changes used Wilcoxon signed‐rank tests. Ninety‐five percent confidence intervals (CIs) for mean differences were computed using bootstrapping (10,000 iterations). Cohen's d for between‐group comparisons was computed as the mean difference divided by pooled standard deviation; Cohen's d for within‐group changes was computed as the mean change divided by the standard deviation of change. Cohen's d effect sizes: 0.2 (small), 0.5 (medium), 0.8 (large). Brain age gap (BAG) by group and timepoint. Note: Values represent mean ± SD of BAG at each timepoint for cross‐sectional comparisons and mean ± SD of BAG change for longitudinal comparisons. The mean difference column indicates the between‐group difference in BAG for cross‐sectional comparisons, mean BAG change for within‐group comparisons, and the between‐group difference in BAG change for the longitudinal between‐group comparison. Cross‐sectional group comparisons and the between‐group difference in longitudinal change used Mann–Whitney U tests; within‐group longitudinal changes used Wilcoxon signed‐rank tests. Ninety‐five percent confidence intervals (CIs) for mean differences were computed using bootstrapping (10,000 iterations). Cohen's d for between‐group comparisons was computed as the mean difference divided by pooled standard deviation; Cohen's d for within‐group changes was computed as the mean change divided by the standard deviation of change. Cohen's d effect sizes: 0.2 (small), 0.5 (medium), 0.8 (large). Brain age gap (BAG) distribution across groups and timepoints. Probability density distributions of BAG values for postpartum and control groups. (A) Baseline: Postpartum women (n = 25, 2.5 ± 2.7 years) vs. controls (n = 30, 0.3 ± 3.6 years), showing a rightward shift in the postpartum group. (B) Follow‐up: Postpartum women (n = 11, 0.1 ± 2.9 years) vs. controls (n = 20, 0.1 ± 3.7 years), showing substantially overlapping distributions. See Figure 1 for statistical comparisons. Analysis of participants with complete data at both timepoints (n = 31; 11 postpartum and 20 nulliparous women) revealed differing longitudinal BAG changes within each group (Figure 1C). Postpartum women showed significant BAG decrease from baseline to follow‐up (−3.3 ± 3.5 years, 95% CI: −5.4 to −1.5 years; Wilcoxon signed‐rank test: z = −2.67, p = 0.008, d = −0.94), while controls showed minimal change (−0.7 ± 4.9 years, 95% CI: −2.8 to 1.5 years; z = −0.90, p = 0.37, d = −0.14). The between‐group difference in BAG change magnitude showed a medium effect size but did not reach statistical significance, with a mean difference of 2.6 years (95% CI: −5.6 to 0.3 years; Mann–Whitney U test: z = 1.55, p = 0.12, d = −0.58). A supplementary linear mixed‐effects model yielded a non‐significant Group × Time interaction in the expected direction (β = −2.6 years, 95% CI: −6.0 to 0.8 years, t = −1.54, p = 0.13; Table S1). Network‐Level SHAP Contributions to BAG Spatial visualization of SHAP values revealed that contributions to brain age predictions were predominantly localized to deep brain structures, including subcortical gray matter, periventricular white matter, and ventricular cerebrospinal fluid (Figure S3). Network‐level aggregation of gray matter SHAP contributions revealed that the subcortical network showed significantly elevated values in postpartum women compared with controls at baseline (Mann–Whitney U test: z = −2.75, p = 0.006, Cohen's d = 0.82), representing the most pronounced group difference among all eight functional networks (Figure 3). This subcortical elevation attenuated at follow‐up, with postpartum and control groups showing similar subcortical contributions (z = −0.54, p = 0.59, d = 0.17). Other functional networks showed no significant group differences at either timepoint. Functional network contributions of Shapley additive explanations (SHAP) to brain age prediction. Mean signed SHAP values (years) for eight functional networks at baseline and follow‐up. Networks: Visual, somatomotor, dorsal attention (DAN), ventral attention (VAN), limbic, frontoparietal (FPN), default mode (DMN), and subcortical. At baseline, the subcortical network showed significantly elevated values in postpartum women (0.48 ± 0.35 years) compared with controls (0.12 ± 0.28 years; Mann–Whitney U test: z = −2.75, p = 0.006, Cohen's d = 0.82). By follow‐up, this difference resolved (postpartum: 0.18 ± 0.22 years; control: 0.14 ± 0.25 years; z = −0.54, p = 0.59, d = 0.17). No other networks showed significant group differences at either timepoint. Error bars indicate standard deviation. *p < 0.05. To further characterize network‐level SHAP contributions, we examined mean absolute SHAP values, which provide a magnitude‐based measure of regional importance independent of directional sign. Across tissue types, no significant between‐group differences in total absolute SHAP magnitude were observed at either timepoint (Table S2), indicating that the postpartum group did not show a generalized elevation in overall SHAP magnitude relative to controls. Within gray matter, mean absolute SHAP proportional contributions showed no significant between‐group differences in any of the eight networks at either timepoint (Table S3), indicating that the significant subcortical group difference identified in the signed SHAP analysis reflects a directional rather than a magnitude difference in network‐level feature importance. Together, these findings suggest that the postpartum elevation in subcortical SHAP contributions represents a selective shift in the direction of subcortical contributions to brain age predictions rather than a broad increase in overall feature importance. A sensitivity analysis recomputing SHAP values using nulliparous controls as the reference distribution yielded consistent conclusions, with the subcortical network remaining the only network showing a significant group difference at baseline (Figure S4). Under the controls‐only reference, control group SHAP values approach zero, such that postpartum values more directly reflect deviations from a normative pattern, further supporting the specificity of subcortical involvement in postpartum BAG elevation. Preliminary Clinical Associations of BAG Across the pooled sample of postpartum and nulliparous women, baseline BAG showed significant correlations with all three primary clinical measures (Figure 4, Table S4): digit span sequencing (Spearman rank correlation: ρ = −0.30, p = 0.03), BDI‐II (ρ = 0.38, p = 0.004), and subjective cognitive complaints (ρ = 0.28, p = 0.04). Within the postpartum group, BAG showed a nominally significant correlation with BDI‐II (ρ = 0.41, p = 0.047). Among secondary clinical measures, BAG significantly correlated with verbal fluency (ρ = −0.33, p = 0.02) in the pooled sample but not with trail‐making performance or estradiol levels. Associations between brain age gap (BAG) and clinical measures at baseline. Scatterplots showing associations between baseline BAG and clinical measures across all participants. (A) Digit span sequencing (n = 54): ρ = −0.30, p = 0.03. (B) Beck Depression Inventory‐Second Edition (BDI‐II): ρ = 0.38, p = 0.004 across all participants (n = 54); ρ = 0.41, p = 0.047 within the postpartum group (n = 24). (C) Subjective cognitive complaints (n = 55): ρ = 0.28, p = 0.04. Lines represent ordinary least‐squares regression fits. Baseline BAG did not differ significantly between APOE E4 carriers and non‐carriers (z = 1.01, p = 0.31, Cohen's d = 0.48), though carriers showed numerically higher values than non‐carriers (carriers: 2.7 ± 4.0 years vs. non‐carriers: 1.1 ± 3.3 years). Subgroup analyses within the postpartum and nulliparous control groups similarly showed no significant associations between APOE status and BAG (Table S5). Discussion This longitudinal study examined whether postpartum women show a biphasic pattern of transient BAG elevation followed by normalization, relative to nulliparous controls. Postpartum women showed significantly elevated BAG at approximately 3 months postpartum, with no significant group difference remaining by approximately 12 months postpartum; however, the formal between‐group comparison of longitudinal change did not reach statistical significance, and the null hypothesis of no differential trajectory between groups could not be rejected. Explainability analysis identified subcortical structures as the network with the most pronounced group difference in SHAP contributions at baseline; this group difference attenuated by follow‐up, paralleling the cross‐sectional BAG pattern. Additionally, associations of BAG with working memory, depressive symptoms, and subjective cognitive complaints provided preliminary evidence that BAG captures clinically meaningful neurobiological variation during the postpartum period, though these correlations were computed across the pooled sample of postpartum and nulliparous women and should be interpreted as exploratory. Brain Age Gap Elevation, Normalization, and Subcortical Plasticity The 2.5‐year BAG elevation at 3 months postpartum, with no significant group difference remaining by 12 months, is descriptively consistent with a transient pattern reflecting the reversible nature of maternal brain changes documented in longitudinal neuroimaging studies (Hoekzema et al. 2017; Servin‐Barthet et al. 2025). The current study extends this evidence by providing a summary metric capturing distributed structural change across the brain in a controlled design. The present findings complement those of Luders et al. (2018), who reported within‐subject BAG decrease from 1–2 days to 4–6 weeks postpartum without a nulliparous control group. As noted, their finding reflects recovery relative to the immediate postpartum nadir rather than BAG relative to nulliparous peers. The present controlled design extends this evidence by showing that postpartum brain structure remains measurably different from that of nulliparous controls at approximately 3 months, with BAG values converging to a statistically indistinguishable level by approximately the first postpartum year. This interpretation is supported by controlled structural studies demonstrating that gray matter volume differences between postpartum women and nulliparous controls persist through at least 3–4 months postpartum (Lisofsky et al. 2019; Nehls et al. 2024), consistent with the elevated BAG observed at approximately 3 months in the present study. Together, these observations suggest a trajectory in which within‐subject structural recovery begins early in the postpartum period, while full convergence to nulliparous‐equivalent levels requires longer, occurring within approximately the first postpartum year. The precise timing of this normalization and its individual variability warrant further investigation in studies with denser longitudinal sampling. The explainability analysis identified the subcortical network as the primary regional contributor to the between‐group BAG difference at baseline, with its SHAP contributions selectively elevated in postpartum women relative to nulliparous controls. This pattern is consistent with documented plasticity in these regions, including the hippocampus, amygdala, basal ganglia, and thalamus, during the early postpartum period (Kim et al. 2010; Luders et al. 2020; Pawluski et al. 2017; Martínez‐García, Paternina‐Die, Desco, et al. 2021). These subcortical regions are implicated in circuits supporting maternal behavior, emotional regulation, and stress response (Duan et al. 2017; Horáková et al. 2022), suggesting that their differential contribution to brain age predictions may reflect adaptive neuroplastic reorganization associated with the demands of the maternal transition, though direct characterization of the underlying structural changes warrants dedicated future investigation. The attenuation of subcortical group differences by 12 months postpartum, concurrent with the normalization of overall BAG values, is consistent with this interpretation. The absence of significant group differences in cortical network SHAP contributions at either timepoint suggests relative specificity to subcortical regions, though the modest sample size limits the detection of smaller cortical effects. Alongside this adaptive interpretation, an important interpretational consideration concerns the relationship between the pretrained model's learned feature representations and postpartum brain changes. Structural brain aging in early adulthood involves changes in subcortical and related deep brain structures (Bethlehem et al. 2022), patterns that may constitute important age‐predictive features in the pretrained model. Prior neuroimaging studies document that pregnancy and the postpartum period are accompanied by substantial structural changes in similar anatomical regions (Hoekzema et al. 2017; Pritschet et al. 2024), raising the possibility that the model's age‐predictive feature space and postpartum neuroplastic changes share overlapping anatomical substrates. The consistent registration of postpartum brains as older‐appearing by the model may therefore reflect, at least in part, this structural overlap rather than accelerated aging per se. We therefore interpret elevated BAG as reflecting a structural deviation within the model's age‐predictive feature space that may partly share anatomical substrates with postpartum neuroplasticity. The precise nature and direction of the structural features underlying this overlap cannot be fully resolved within the present BAG‐based framework, and elucidating the relationship between postpartum neuroplastic changes and normative aging trajectories in early adulthood represents an open empirical question for future research. Convergent Evidence From Biological Aging and Long‐Term Implications The transient brain age advancement observed at 3 months postpartum may reflect broader cellular aging processes documented during pregnancy. Previous research demonstrated associations between pregnancy and cellular aging markers, including telomere shortening (Pollack et al. 2018) and epigenetic age acceleration (Ryan et al. 2018). Remarkably, recent longitudinal studies documented recovery patterns in cellular aging markers during postpartum, with epigenetic aging measures showing reversal over the first postpartum year (Ross et al. 2020). This convergent evidence across multiple biological aging metrics suggests that structural brain aging and cellular aging processes may exhibit similar trajectories of initial elevation followed by recovery, reflecting coordinated biological adaptations to the dramatic physiological transitions of pregnancy and childbirth. The normalization we observed at 12 months postpartum may represent an intermediate step toward a more lasting structural reorganization. Cross‐sectional population studies have reported lower BAG in parous compared with nulliparous women, with effects detectable decades after childbirth (de Lange et al. 2019), raising the intriguing possibility that the maternal brain may progress beyond mere recovery toward a state of enhanced biological resilience. Whether this trajectory continues beyond the first postpartum year, and what mechanisms underlie such long‐term changes, remains important topics for future longitudinal research. Implications for Understanding Maternal Brain Health The findings may help explain heterogeneity in previous maternal brain research. The potential biphasic pattern we observed suggests that studies conducted at different postpartum timepoints may capture different phases of this dynamic process, potentially contributing to inconsistent findings across studies. The normalization of BAG by 12 months is consistent with the temporary nature of pregnancy‐related brain changes and may help explain why single‐timepoint studies yield conflicting results. Understanding this temporal trajectory may inform the design of future studies and the interpretation of their results within the broader context of maternal brain adaptation. The selective association of subcortical structures with postpartum BAG elevation suggests that these regions merit particular attention in future investigations of postpartum neuroplasticity. Future studies could investigate factors that moderate recovery trajectories, such as sleep quality, social support, and maternity leave policies (Dagher et al. 2014). Additionally, BAG may provide a useful quantitative index for future studies examining factors that influence maternal brain health during the postpartum period. Preliminary Clinical Associations of BAG Across the pooled sample of postpartum and nulliparous women, baseline BAG showed associations with working memory, depressive symptoms, and subjective cognitive complaints, consistent with broader literature linking elevated brain aging to cognitive and psychiatric outcomes (Cole et al. 2018; Kaufmann et al. 2019). However, these correlations require cautious interpretation: computed across the pooled sample, they may partly reflect between‐group mean differences rather than within‐postpartum brain–behavior relationships. Within the postpartum group alone, only the BAG–BDI‐II association reached statistical significance, limiting inference about postpartum‐specific brain–behavior coupling. The nominally stronger association between BAG and depressive symptoms within the postpartum group relative to the overall sample is consistent with prior studies linking postpartum brain changes to emotional symptom vulnerability (Gingnell et al. 2015; Wonch et al. 2016). These pooled associations are therefore interpreted as preliminary, exploratory evidence pending replication in larger within‐group designs. The association between BAG and subjective cognitive complaints in the pooled sample provides preliminary support for a neurobiological basis of these frequently reported experiences during the postpartum period. The temporary nature of elevated BAG observed in the present study, together with prior evidence suggesting that postpartum cognitive complaints tend to resolve over time (Orchard et al. 2022), is consistent with the interpretation that these experiences reflect transient neuroplastic adaptation rather than permanent impairment. Whether BAG normalization and symptomatic recovery follow parallel trajectories within individuals remains to be established in studies with repeated clinical assessments. Notably, BAG showed no significant correlation with concurrent estradiol levels despite clear group differences in hormone concentrations. Estradiol is a well‐established driver of pregnancy‐related brain structural changes (Hoekzema et al. 2017; Servin‐Barthet et al. 2025), yet a single hormone measurement at 3 months postpartum may not adequately capture the cumulative gestational hormonal exposure that shaped the structural reorganization indexed by BAG. Estradiol recovery during the postpartum period is further complicated by breastfeeding status, which prolongs estrogen suppression and introduces substantial inter‐individual variability at any given postpartum timepoint. The null estradiol–BAG finding therefore does not preclude a hormonal basis for postpartum brain aging patterns; rather, it highlights the methodological challenge of capturing cumulative hormonal influences through a single cross‐sectional measurement. Limitations and Future Directions Several limitations warrant consideration when interpreting the findings. First, the modest sample size, particularly for longitudinal analysis, limits statistical power for detecting subtle effects and precludes subgroup analyses exploring potential moderators of recovery trajectories. The between‐group difference in longitudinal BAG change showed a medium effect size but did not reach statistical significance. While the non‐significant result should not be attributed specifically to limited power, as such an interpretation cannot be distinguished from a true absence of a group difference, future studies with larger cohorts are needed to evaluate group differences in recovery trajectories and identify predictors of individual variation. Second, generalizability is limited by the relatively homogeneous sample characteristics. All participants were East Asian women, highly educated, recruited from a single medical center. Whether similar biphasic BAG trajectories occur in ethnically and culturally diverse populations, in women with pregnancy or delivery complications, or across different socioeconomic contexts remains to be established. Cross‐cultural, multinational studies with diverse sampling are needed to assess the broader applicability of the findings. Third, clinical correlations were limited to baseline assessments, preventing examination of longitudinal brain‐behavior relationships. Additionally, although breastfeeding status was recorded at assessment, the available information was too limited to adequately characterize the influence of breastfeeding on postpartum estrogen suppression and hormonal recovery, which may represent an important moderator of both BAG patterns and estradiol–BAG associations. Future studies should incorporate more comprehensive documentation of breastfeeding practices to enable precise examination of hormonal moderators. Fourth, the timing of the baseline assessment may have missed the peak of brain age advancement, as intensive longitudinal studies demonstrate maximal structural changes occur during late pregnancy and immediate postpartum (Pritschet et al. 2024; Nehls et al. 2024). Studies with denser temporal sampling beginning immediately after childbirth would better characterize the timing of peak brain age advancement and the onset of recovery. Finally, the brain age prediction model relied solely on structural MRI data. Future research may benefit from multimodal approaches incorporating functional connectivity, white matter microstructure, and metabolic markers (Hoekzema et al. 2022), which may provide more comprehensive characterization of maternal brain changes and improve sensitivity to subtle postpartum effects. Conclusions This study provides evidence of transient BAG elevation in postpartum women relative to nulliparous controls at approximately 3 months, with no significant group difference remaining by 12 months. Whether this pattern reflects a trajectory genuinely distinct from that of nulliparous women remains to be established in larger longitudinal samples. These findings are nonetheless consistent with an adaptive neuroplasticity framework, suggesting that postpartum brain structural change may reflect a temporary state rather than detrimental aging. BAG offers a quantitative summary index of distributed structural change across the brain, complementing approaches that examine specific brain regions. Replication in more diverse cohorts is needed to establish the generalizability of these findings and their relevance to maternal brain health across different populations and contexts. Funding This work was supported by the National Research Foundation of Korea (NRF) funded by the Korean government (RS‐2023‐00240457 to C.P.; NRF‐2014R1A1A2055116 and RS‐2025‐00562956 to N.‐Y. S.). Conflicts of Interest The authors declare no conflicts of interest. Supporting information Data S1: Supporting Information. Table S1: Linear mixed‐effects model for brain age gap longitudinal trajectories. Table S2: Total absolute SHAP magnitude by tissue type. Table S3: Mean absolute SHAP proportional contributions by functional network. Table S4: Association between brain age gap and clinical measures. Table S5: Brain age gap by APOE E4 carrier status at baseline. Figure S1: Brain age prediction model architecture. Figure S2: Chronological versus predicted brain age by group. Figure S3: Tissue‐specific SHAP value spatial distributions. Figure S4: Sensitivity analysis of network‐level SHAP contributions. Data Availability Statement The Human Connectome Project data used for brain age prediction model pretraining are publicly available at https://db.humanconnectome.org. Individual participant data from the postpartum cohort cannot be made publicly available due to ethical restrictions and privacy concerns related to identifiable human subjects. The data are available from the corresponding author upon reasonable request. Code for the brain age prediction model is available at https://gitlab.com/chang‐hyun.park/postpartum‐brain‐age‐prediction. References Associated Data Supplementary Materials Data S1: Supporting Information. Table S1: Linear mixed‐effects model for brain age gap longitudinal trajectories. Table S2: Total absolute SHAP magnitude by tissue type. Table S3: Mean absolute SHAP proportional contributions by functional network. Table S4: Association between brain age gap and clinical measures. Table S5: Brain age gap by APOE E4 carrier status at baseline. Figure S1: Brain age prediction model architecture. Figure S2: Chronological versus predicted brain age by group. Figure S3: Tissue‐specific SHAP value spatial distributions. Figure S4: Sensitivity analysis of network‐level SHAP contributions. Data Availability Statement The Human Connectome Project data used for brain age prediction model pretraining are publicly available at https://db.humanconnectome.org. Individual participant data from the postpartum cohort cannot be made publicly available due to ethical restrictions and privacy concerns related to identifiable human subjects. The data are available from the corresponding author upon reasonable request. Code for the brain age prediction model is available at https://gitlab.com/chang‐hyun.park/postpartum‐brain‐age‐prediction.