We obtained data from the Longitudinal Cognitive Impairment and Dementia Study (DELCODE) study, conducted by the German Center for Neurodegenerative Diseases (DZNE) (Jessen et al., 2018). This is an ongoing German multicenter observational study on predementia AD that was initiated in 2014. It aims to characterize early disease stages, in particular SCD, improve prognostics of disease progression and identify new markers for preclinical AD.¹

All participants or their representatives gave written informed consent. The study protocol was approved by the local institutional review boards and ethics committees of the participating centers. The study was conducted in accordance with the Declaration of Helsinki of 1975 and its subsequent amendments.

From the DELCODE cohort, we selected participants with MRI scans available at baseline and at least one follow-up (Figure 1). Diagnosis was extracted at baseline and treated as a time-invariant factor to examine how initial clinical phenotype relates to subsequent longitudinal brain changes. Subjects were grouped according to the following diagnostic groups (Cognitively Normal [CN], Subjective Cognitive Decline [SCD], Mild Cognitive Impairment [MCI], and Alzheimer’s Disease [AD]), and patients ‘relatives. The term SCD was defined as a persistent self-perceived cognitive decline in the absence of objective cognitive impairment, as measured by the cognitive test battery Consortium to Establish a Registry for Alzheimer’s Disease (CERAD). This self-perceived cognitive decline was required to be lasting at least 6 months and be unrelated to an acute event (Jessen et al., 2014). The MCI patients were diagnosed according to the core clinical criteria for MCI of the National Institute on Aging-Alzheimer’s Association (NIA-AA) workgroup guidelines (Albert et al., 2011). The CN participants exhibited no objective cognitive impairment in cognitive tests, no history of neurological or psychiatric disease, and no self-reported cognitive decline. The term “relatives” was defined as individuals who were cognitively normal and had at least one individual with confirmed AD in their immediate family. The DELCODE participants are subject to annual follow-ups, during which clinical, neuropsychological, and imaging assessments are conducted. The sample size was not based on a priori power calculation but rather on the current data availability.

In this study, we extracted some tests from the CERAD battery which is carried out as part of the DELCODE routine. We used the Mini-Mental State Examination (MMSE), the Digit Span Test total score (forward and backward) from the Wechsler Memory Scale-Revised (WMS-R) to evaluate working memory. Furthermore, we employed the ratio of the Trail Making Test B to A (TMTB/A) as a second measure of executive function.

Cerebrospinal fluid (CSF) was obtained by means of lumbar puncture following internal consensus recommendations (Jessen et al., 2018). Biomaterial sampling procedures and biomarker measurements for DELCODE have been described elsewhere (Jessen et al., 2018). Amyloid positivity was determined using CSF Aß42/Aß40 ratio and tau pathology by using total tau and phospho tau 181 levels. The cut-off for abnormal concentrations of Aß42/Aß40 ratio was derived from the literature, which applied the widely used threshold of 0.1, commonly employed as a biomarker criterion for Alzheimer’s disease (Janelidze et al., 2016). In contrast, cut-offs for tau biomarkers followed a univariate two-component Gaussian mixture modeling (Colin et al., 2023; Quattrini et al., 2023) fitted to the full set of available baseline values, without sub-group stratification with the midpoint between the means used as a threshold. For total tau, the resulting cutoff was 580 pg./mL, and for phosphorylated tau 181, the cutoff was 76 pg./mL (Supplementary materials, Section1.5–1.6). To control for coherency, we also performed a sensitivity analysis, using the cut-offs predefined in the original DELCODE paper. After grouping individuals based on these cut-offs, we binary coded tau and Aß variables into positive and negative groups.

The MRI data were acquired from nine Siemens 3.0 Tesla MRI scanners (four Verio, one Skyra, three TimTrio, and one Prisma system) using identical acquisition parameters and instructions (Jessen et al., 2018). This is a highly standardized, centrally controlled protocol with harmonized sequences, centralized training, traveling-head qualification, and ACR phantom monitoring across all nine Siemens sites (Jessen et al., 2018). Additionally, to ensure the consistency of image quality throughout the acquisition phase, all scans were subjected to a semi-automated quality check during the study’s conduction. This protocol deviation reporting mechanism enabled the communication of any issues to the respective study sites, facilitating the adjustment of the acquisition at the site. Such procedures of first-level quality checks (e.g., motion artifacts correction) were performed centrally by DZNE Bonn and Magdeburg. After we acquired the data, we performed again a data quality check with default parameters in CAT12 combined with visual inspection (FL), resulting in a total MRI drop-out rate relative to the full acquired dataset of 0.26%. High-resolution structural images were obtained using a sagittal T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) sequence (field of view = 256 × 256 mm²; matrix = 256 × 256; isotropic voxel size = 1 mm³; repetition time = 2,500 ms; echo time = 4.37 ms; flip angle = 7°; 192 slices; parallel imaging acceleration factor = 2).

We used distinct preprocessing pipelines for analyzing baseline and longitudinal MRI data.

Post-hoc, we conducted a series of sensitivity analysis. To control for multiple testing within our a priori hypotheses, we applied Holm’s correction within pre-specified families of tests. Specifically, sex main effects and sex × time interactions were evaluated using Type III tests across both regions of interest, and ApoE ε4 baseline effects were evaluated using planned contrasts of estimated marginal means comparing heterozygotes versus non-carriers and homozygotes versus heterozygotes in each region. Other terms in the models, including diagnosis, pathology subgroup variables, and higher-order interaction terms, were included as covariates or exploratory effects and were not subjected to additional multiplicity correction, consistent with standard practice in mixed-effects regression. For all fixed effects, we report unstandardized regression coefficients with 95% confidence intervals as effect size estimates, along with exact p-values. To assess whether the null effect of amyloid/tau status on baseline hippocampal volume reflects sufficient statistical power, we calculated the minimum detectable effect (MDE).

In addition, we quantified the magnitude of ApoE ε4 –related group differences using standardized effect sizes (Cohen’s d). Cohen’s d was calculated for planned ApoE ε4 contrasts in basal forebrain and hippocampus using estimated marginal means derived from the mixed-effects models. Effect sizes were computed using the residual standard deviation of each model, with degrees of freedom inherited from the Kenward–Roger approximation. Regarding TIV method’s correction, we also conducted a sensitivity analysis using residual-based TIV correction after regressing out sex for both basal forebrain and hippocampus.

The chi-square test for sex differences showed that the number of males (n = 349) and females (n = 363) was balanced (Table 1). The results of the Wilcoxon rank-sum test indicated a statistically significant difference in TIV at baseline between male (mean = 1578.691) and female (mean = 1379.152) participants (W = 2,617,555, p < 2.2 × 10⁻¹⁶). A highly significant difference was observed in the distribution of ApoE ε4 status, indicating a relatively small proportion of homozygous carriers. Among the ApoE ε4 positive both homozygote and heterozygote (n = 236), 119 (50.04%) were women and 117 (49.57%) were men. Participants were divided in the following diagnostic groups: CN (N = 184), SCD (N = 331), MCI (N = 128) and AD (N = 69). The overall median age of the sample was 71.25 years (interquartile range [IQR] = 9.22). The median number of years of education was 14.00 (IQR = 5), suggesting a relatively well-educated cohort. Lastly, the median duration of follow-up across the sample was 2.8 years [IQR] = 1.75. More detailed demographic characteristics and statistical tests are summarized in Table 1. Among the analyzed 712 MRI scans with repeated measures, a subset of 344 individuals had baseline CSF data, specifically 161 individuals were amyloid-positive (Aß 42/40 ratio) and 183 were amyloid-negative. For phosphorylated tau (p-tau181), 75 individuals were classified as positive and 269 as negative at baseline. In terms of total tau (t-tau), 81 individuals were positive and 263 were negative at baseline.

After Holm correction, the main effect of sex remained significant for both basal forebrain and hippocampal volumes. The interaction of sex with time stayed not significant in both region after correction. Regarding ApoE ε4, heterozygous carriers showed significantly lower basal forebrain and hippocampal volumes compared with non-carriers, and these effects survived Holm correction. In the basal forebrain, ApoE ε4 heterozygotes showed moderately lower volume compared with non-carriers (Cohen’s d = −0.50, 95% CI [−0.91, −0.09]), whereas the difference between ApoE ε4 homozygotes and heterozygotes was negligible and not statistically meaningful (d = −0.06, 95% CI [−0.58, 0.47]). After correction, no longitudinal ApoE ε4 -related differences in basal forebrain atrophy rates were detected. In the hippocampus, ApoE ε4 heterozygotes exhibited substantially lower volume compared with non-carriers, corresponding to a large effect size (d = −1.76, 95% CI [−2.72, −0.80]). ApoE ε4 homozygotes showed an even larger reduction in hippocampal volume relative to heterozygotes (d = −4.68, 95% CI [−5.91, −3.44]), indicating a strong dependent effect of ApoE ε4 on hippocampal structure. In addition, ApoE ε4 homozygotes exhibited significantly lower hippocampal volume compared with heterozygotes, an effect that also remained significant after correction. All corrected and uncorrected results are summarized in Table 3.

To assess whether the null effect of amyloid/tau status on baseline hippocampal volume reflects sufficient statistical power, we calculated the minimum detectable effect (MDE). Based on the observed standard error (SE = 0.0285) and degrees of freedom (df = 344), the MDE was 0.056. Thus, the present sample was only powered to detect effects of moderate magnitude.

The results of the residual-based TIV correction using sex as regressor were highly consistent with the primary analyses (Supplementary Table S12). Specifically, diagnostic group and age effects remained robust, ApoE ε4 status continued to be associated with smaller hippocampal volumes, and critically, no additional sex × time or ApoE ε4 × time effects emerged in the basal forebrain under residual-based correction. In the hippocampus, faster longitudinal atrophy in ApoE ε4 homozygotes remained evident.

At baseline, we found several demographic and genetic factors, which showed small but noteworthy associations with basal forebrain structure. Contrary to our hypothesis, women exhibited larger basal forebrain volumes than men after TIV proportion adjustment method, whereas the direction reversed in non TIV-adjusted analyses, reflecting expected sex differences in overall head size. Similarly, ApoE ε4 status demonstrated only a subtle influence on basal forebrain volume, which contradicts our initial assumption. Indeed, ApoE ε4 heterozygotes showed marginally smaller volumes than non-carriers, but no differences emerged between ε4 heterozygotes and homozygotes. Age effects were as expected, with older participants displaying smaller basal forebrain volumes. Diagnostic comparisons indicated that individuals with subjective cognitive decline did not yet show basal forebrain atrophy, consistent with the notion that early, preclinical complaints are not accompanied by detectable cholinergic degeneration. In contrast, participants with MCI and especially AD exhibited substantially reduced basal forebrain volumes. Neuroinflammation biomarkers provided a similar picture, as lower amyloid burden was associated with larger basal forebrain volume, while total tau and p-tau levels were not significantly related to baseline structure. Together, these findings suggest that basal forebrain alterations at baseline remain subtle in preclinical groups and become more pronounced with advancing clinical stage. However, our findings on baseline sex differences favoring larger basal forebrain and hippocampal volumes in women are not consistent with previous studies showing women higher susceptibility to neurodegenerative processes or have a greater pathological burden (Gamache et al., 2020). On the other hand, they are consistent with the existing literature showing greater susceptibility in men (Cavedo et al., 2018; Neu et al., 2017; Snyder et al., 2016). Men generally have higher cardiovascular risk factors, which can elevate the probability of developing vascular dementia, while women with hypertension appear to have a disproportionately higher risk of developing dementia and exhibit poorer cognitive performance compared to normotensive women (Gamache et al., 2020). This highlights the importance of lifestyle a potential influence in cognition affecting brain volumes differently in men and women. Since our analyses did not include cardiovascular measurements, the interpretation of potential underlying biological mechanisms remains speculative. Therefore, this discussion should be considered as hypothesis generating rather than conclusive, given the observational nature of the data. Nonetheless, we cannot exclude the possibility that men in our sample had higher baseline levels of cardiovascular risk factors, which may have negatively affected brain volumes. Previous research has shown that men are more likely to exhibit signs of neurodegeneration, particularly in the hippocampus (Jack et al., 2017), which partly aligns with our results. Another study has shown that sleep deficit is associated with negative emotions and lower basal nucleus of Meynert functional connectivity with the posterior cingulate cortex in men (Li et al., 2023).

The association between lower basal forebrain volume and poorer working memory performance, but not executive function as assessed by the TMT B/A ratio, suggests that cholinergic degeneration may preferentially affect attentional and short-term memory processes. This interpretation is consistent with experimental and clinical evidence highlighting the critical role of basal forebrain cholinergic projections in sustaining working memory and attentional modulation (Sarter et al., 2003; Mesulam, 2013). In contrast, executive functions rely on distributed frontoparietal networks and multiple neurotransmitter systems, which may explain the absence of a direct association with basal forebrain volume alone (Alvarez and Emory, 2006; Lezak et al., 2012). Similarly, despite well-documented sex differences in cognitive aging and AD risk (Cavedo et al., 2018; Gamache et al., 2020), no significant sex-related differences in cognitive performance were observed in this analysis. These findings may indicate that other factors, such as baseline cognitive reserve, education, or lifestyle variables, may play a more substantial role in determining individual cognitive trajectories.

Contrary to our initial assumptions, we found that sex as well as ApoE ε4 status were not associated with basal forebrain rate of atrophy over time. However, this was partly true for the hippocampus, where being a carrier of the ApoE ε4 (heterozygote) was associated with greater decline over time. Nonetheless, sex did not appear to be a predictive factor for brain atrophy in either region. Our findings also showed that hippocampus, but not basal forebrain, exhibited significant progressive atrophy over time. In both regions, atrophy was associated with baseline diagnosis. Previous work suggests that basal forebrain degeneration may occur early in the AD continuum but progress relatively slowly, whereas hippocampal atrophy accelerates at later stages and is more readily detectable over short follow-up intervals (Grothe et al., 2012; Schmitz and Spreng, 2016; Mesulam, 2013). This stage-dependent pattern may explain why hippocampal, but not basal forebrain, atrophy was observed longitudinally in the present study. Longitudinal results showed that amyloid-negative individuals with SCD and MCI exhibited significant slower basal forebrain atrophy over time compared to amyloid-positive individuals with SCD and MCI. The fact that basal forebrain atrophy was detected only in preclinical and prodromal stages of AD is in line with previous literature (Bohnen et al., 2018; Chiesa et al., 2019; Xia et al., 2024). Indeed, Daamen and colleagues found reduced basal forebrain volume in amyloid-positive individuals with SCD compared to amyloid-negative individuals with SCD, suggesting that once amyloid pathology is accompanied by cognitive symptoms, cholinergic degeneration becomes detectable (Daamen et al., 2023). Moreover, a recent study showed evidence that lower basal forebrain volume was associated with faster amyloid-beta accumulation but that, amyloid-beta pathology alone, was not associated with faster basal forebrain atrophy (Yoo et al., 2024). This might suggest that amyloid pathology by itself is insufficient to drive basal forebrain atrophy and that measurable decline appears only when amyloid positivity is accompanied by early cognitive symptoms. This interpretation is also consistent with recent longitudinal work in cognitively unimpaired older adults, where individuals progressing from amyloid negative to amyloid positive pathology showed no accelerated volume loss in the basal forebrain or hippocampus compared to controls (Xia et al., 2024).

In our analysis, tau pathology did not predict longitudinal basal forebrain or hippocampal volumes, nor did it interact with sex or ApoE ε4 status. Instead, hippocampal volume declined significantly faster over time in tau-positive individuals with MCI and AD compared to tau-negative cognitively normal individuals. Nonetheless, our findings are in contrast with evidence increasingly supporting the influence of tau on the cholinergic system (Berry and Harrison, 2023). Indeed, a longitudinal study showed tau accumulation as the strongest predictor of basal forebrain atrophy over time (Yoo et al., 2024). Supporting this claim a recent imaging study showed evidence that the basal forebrain cholinergic system degenerates in individuals at risk for AD presenting tau positive markers (Cantero et al., 2020). The observed differences in our results compared to the literature may be attributable to differences in methodology used. Specifically, the latter study utilized amyloid and tau-PET imaging (Cantero et al., 2020), whereas the former was a cross-sectional analysis from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort with a relatively small sample size (Yoo et al., 2024). In the ADNI study, cytoarchitectonic probabilistic maps of the basal forebrain magnocellular compartments derived from 10 subjects (Záborszky et al., 2018) were employed, with extensions mapped to the amygdala, hippocampus, and neocortex (Yoo et al., 2024). These differences suggest that amyloid-related basal forebrain vulnerability may be more detectable in preclinical and early symptomatic stages, whereas tau-related effects may predominate at later disease stages. Consistently with our findings, a recent cross-sectional study analyzing autopsy data showed that, within the AD subtypes, there seem to be different patterns of basal forebrain vulnerability depending on age, ApoE ε4 status and sex (Hanna Al-Shaikh et al., 2020).

Regarding cognition, time was found to be a significant predictor of executive function but not working memory, as evidenced by worsening performance on the TMT over time, while Digit Span scores remained stable. The decline in TMT performance suggests a gradual deterioration of processing speed and cognitive flexibility, functions that are often sensitive to aging and neurodegenerative processes. In contrast, the stability of Digit Span scores may be attributed to practice effects, as repeated exposure to working memory tasks over multiple testing sessions can lead to performance improvements or maintenance, masking underlying cognitive decline (Teipel et al., 2023a,b). This differential pattern of change across cognitive domains underscores the importance of longitudinal assessments in distinguishing true cognitive decline from practice-related improvements.

This study has some limitations. First, the basal forebrain is a small and anatomically complex region, which presents challenges for reliable volumetric measurement using standard MRI segmentation techniques. We relied on a basal forebrain mask that was anatomically defined from stereotactic atlases and functionally validated in independent samples (Fritz et al., 2019). Although it is possible that our basal forebrain masks (Fritz et al., 2019) lacked sufficient sensitivity to detect subtle preclinical changes, we conducted a sensitivity analysis using masks that delineate basal forebrain nuclei derived from post-mortem MRI and histological data (Kilimann et al., 2014). This approach yielded comparable results, suggesting that mask resolution was not a major limiting factor in our study. Nevertheless, differences from previously published results may reflect methodological variations (e.g., parcellation approaches, preprocessing pipelines) or cohort characteristics rather than inaccuracies in the applied mask. Second, CSF biomarkers and MRI-derived volumetric measures may not fully capture the extent or distribution of amyloid and tau pathology, nor do they necessarily reflect the functional integrity of the basal forebrain, particularly its cholinergic projections. More sensitive modalities, such as PET imaging capable of detecting cholinergic synaptic dysfunction or diffusion MRI for assessing microstructural changes in cholinergic white matter tracts, may offer a more accurate assessment of basal forebrain degeneration in vivo. Future studies employing such modalities could provide a more nuanced understanding of BF involvement in early neurodegeneration. Third, the presence of MCI within the diagnostic spectrum may have obscured early, more subtle changes in basal forebrain structure. MCI is associated with widespread brain alterations, potentially masking region-specific effects of early degeneration. However, our sensitivity analysis focusing on individuals with SCD compared to CN participants did not reveal evidence of basal forebrain atrophy, suggesting that basal forebrain degeneration may occur later in the disease trajectory or may be detectable only through more sensitive imaging techniques.

Regarding sex-based differences, we did not analyze cardiovascular risk factors, and hormonal data were not available. These mechanisms remain speculative and should be examined in future analyses incorporating vascular risk measures. Moreover, TIV represents a major confounding factor when estimating local volumes of interest, as it has been shown that different TIV-adjustment methods can reduce the number of sex differences or generate larger adjusted volumes in females, promoting sex differences that are mainly attributable to TIV variation (Sanchis-Segura et al., 2019). Additionally, we performed a sensitivity analysis using TIV-residuals correction and found concordant results. Thus, our main conclusions did not seem to be dependent on the choice of TIV correction method. Nonetheless, our baseline sex differences following TIV correction should be interpreted as relative, method-dependent effects rather than absolute volumetric differences. A further constraint pertains to the comparatively modest estimate magnitude exhibited by ApoE ε4 copies. This characteristic renders the distinctions between genetic variants with a concomitantly low level of robustness. A larger sample of ApoE ε4 homozygous individuals, which is in turn extremely rare, would be necessary to determine whether the trajectory of neurodegeneration differs significantly from that of heterozygotes, homozygotes and non-carriers, or to observe its interaction with sex. Future studies should also continue to explore whether sex-dependent neurobiological mechanisms, including hormonal influences, interact with genetic risk factors such as ApoE ε4 to shape patterns of brain atrophy.

To conclude, our study revealed that baseline diagnosis, but not sex or ApoE ε4 status, predicted basal forebrain atrophy rate. In contrast, hippocampal volume reduction was predicted by ApoE ε4 homozygosis at baseline and showed significant progressive atrophy over time, particularly in individuals with MCI and AD. Tau and amyloid pathology alone were not predictive of basal forebrain atrophy unless clinical symptoms were present, with amyloid, but not tau, already emerging as a sensitive marker of early cholinergic basal forebrain vulnerability. These findings suggest that basal forebrain degeneration may occur independently of early tau pathology, emphasizing the need for interventions that protect cholinergic function across different genetic and sex groups.