Hierarchical disruption of lateral prefrontal cortex gradients in cognitive aging

Abstract

The lateral prefrontal cortex (LPFC) plays a pivotal role in executive functions and exhibits a hierarchical rostro-caudal organization critical for higher-order cognition. Using connectome gradient mapping of resting-state fMRI data across young, middle-aged, and older adults (N = 478), we found preserved global gradient structure but significant compression of the principal gradient in older adults relative to middle-aged adults, particularly in dorsolateral (DLPFC) and frontopolar (FPC) regions. This reduced functional differentiation corresponded to lower spatial separation between LPFC subdivisions. Meta-analytic decoding linked these changes to attenuated engagement of executive functions. Crucially, in an independent cohort of older adults (N = 99), individuals with better executive function exhibited greater gradient range and variation at the global level, along with higher gradient values in the DLPFC and ventrolateral prefrontal cortex (VLPFC) and lower values in the premotor cortex at the regional level. These findings suggest that age-related disruption of LPFC gradient organization may reflect neural dedifferentiation and is closely related to executive decline. Gradient compression in the LPFC may serve as a novel biomarker of cognitive aging, offering insights into the hierarchical reorganization of brain networks in late life.

Introduction

The lateral prefrontal cortex (LPFC) is a critical brain region involved in higher cognitive functions, including executive functions such as working memory, cognitive flexibility, and inhibition, as well as declarative memory, reasoning, creative thinking, and theory of mind [1–5]. Anatomically, the LPFC is characterized by a distinct granular layer [6, 7] and is extensively interconnected with other cortical and subcortical regions vital for cognition [8–10]. Notably, the LPFC undergoes protracted maturation and is among the first regions to exhibit age-related decline, making it a key target in understanding cognitive aging.

Aging reshapes brain connectivity by increasing interhemispheric communication, reducing within-network cohesion, and enhancing between-network interactions. Maintaining a balance between intra- and inter-network integration is crucial for sustaining cognitive function across the lifespan. However, how these changes manifest in the hierarchical organization of the LPFC remains unclear. Due to the complexity of brain connectivity, more refined approaches are needed to study these changes [11]. One promising approach is to examine hierarchical brain organization, a fundamental principle of cortical architecture that supports efficient information encoding and integration [12]. This hierarchy spans from primary sensory areas, responsible for basic perceptual processing, to transmodal association areas, such as the LPFC, which support abstract and integrative cognitive functions [13].

Traditional neuroimaging studies often assume that functional connectivity within specific brain regions is uniform and invariant [14]. However, recent advances in dimensionality reduction techniques allow high-dimensional resting-state functional magnetic resonance imaging (fMRI) data to be mapped into low-dimensional functional connectivity gradients, offering a more refined and continuous representation of brain organization [15, 16]. Unlike traditional region-based analyses, gradient mapping offers a holistic and topographic view of brain organization, capturing continuous shifts in inter-regional connectivity and revealing how different neural systems integrate to support cognition [17, 18]. In this framework, functional gradients capture the principal axes along which connectivity profiles systematically vary across the cortex. These gradients effectively represent a continuous spatial spectrum that indexes the brain’s hierarchical organization—from lower-order sensory and motor regions to higher-order transmodal areas. As such, they offer a quantitative and spatially extended depiction of cortical hierarchy, reflecting how information is progressively integrated and abstracted along cortical processing streams. This gradient-based approach has redefined our understanding of hierarchical patterns in a variety of cortical regions, including the somatosensory cortex, angular gyrus, cingulate cortex, insula, and hippocampus [19–23].

Emerging evidence suggests that cortical gradient organization undergoes specific alterations with aging, although the general pattern was preserved. For instance, older adults exhibit higher gradient values in the right superior parietal lobe and sensorimotor cortex, but lower values in occipital and ventral temporal regions compared to young adults [24]. Furthermore, the gradient axis showed a compression trend that was more pronounced under natural stimulation conditions [25]. Previous studies have well characterized age-related compression of functional gradients at the whole-brain level, reflecting a global dedifferentiation process across sensory, motor, and heteromodal networks [25, 26]. Wang et al. systematically delineated lifespan trajectories of functional gradients and their associations with cognitive domains including working memory and visuospatial ability [27]. In addition, recent meta-analytic evidence has delineated the intrinsic rostro-caudal and dorsoventral functional gradients of the LPFC in healthy adults, highlighting its hierarchical organization related to cognitive control and abstraction [28]. Further, several studies have reported age-related functional reorganization in the prefrontal cortex, including activation shifts and dedifferentiation [29, 30]. However, it remains unclear whether the LPFC—a region critically involved in higher-order cognitive control—undergoes specific, spatially localized gradient alterations during aging, and whether such changes are selectively linked to executive function. Moreover, although large-scale lifespan studies suggest non-linear trajectories of global functional hierarchy [27], it is unknown whether the LPFC follows a similar pattern or exhibits regionally distinct developmental courses.

To address these gaps, the present study specifically investigated LPFC-centric connectivity gradients across three carefully defined age groups (young, middle-aged, and older adults). This design not only allowed us to identify the precise lifespan trajectory of LPFC gradient organization but also to test whether LPFC-specific gradient features are selectively associated with executive performance in later life. Grounded in the dedifferentiation hypothesis of cognitive aging—which posits a decline in functional specialization with age [31–33]—we hypothesized that aging would reduce the distinctiveness of connectivity profiles between LPFC subregions, resulting in more homogeneous gradient patterns. We further predicted that such gradient compression would be linked to declines in higher-order cognitive abilities, particularly those reliant on the prefrontal cortex.

Material and methods

Participants

The data for this study were obtained from the publicly available Southwest University Adult Lifespan Dataset (SALD, https://doi.org/10.15387/fcp_indi.sald) [34]. The dataset includes structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI) data of 494 healthy individuals aged 18–80 years. Full details about the dataset (e.g., ethics approval, informed consent, and inclusion and exclusion criteria) can be found in previous literature [34]. We excluded participants with poor image quality or excessive head motion during scanning (mean absolute framewise displacement > 0.5 mm). The final sample consisted of 478 participants (183 males and 295 females), divided into three age groups: young adult (19–40 years old, n = 195), middle-aged adult (41–59 years old, n = 169), and older adult (60–80 years old, n = 114) (Fig. 1B). No significant differences in gender distribution were observed among the three groups (p = 0.202). To account for potential motion-related confounds, head motion was quantified as mean framewise displacement for each participant. Given the significant differences in mean head motion among the three groups (young adults: 0.13 ± 0.07 mm; middle-aged adults: 0.20 ± 0.09 mm; older adults: 0.24 ± 0.09 mm; F = 69.07, p < 0.01), head motion was included as a covariate in all subsequent analyses.

Fig. 1.

A Lateral prefrontal cortex (LPFC) regions defined via the Human Brainnetome Atlas. PMC: premotor cortex, FEF: frontal eye fields, IFC: inferior frontal cortex, VLPFC: ventrolateral prefrontal cortex, DLPFC: dorsolateral prefrontal cortex, and FPC: frontopolar cortex. B Sample distribution by age group and sex in the SALD dataset. C Left: Average explanation ratios of the first 20 principal gradient components for each age group. Right: Cumulative explanation ratios of the first 20 principal gradient components for each age group

Given the absence of behavioral data in SALD, we independently collected sMRI, rs-fMRI, and executive function data from 100 healthy older adults. The inclusion criteria for participants were as follows: (1) age ≥ 60 years; (2) years of education ≥ 6 years; (3) Mini-Mental State Examination (MMSE) score ≥ 24; (4) normal hearing and vision, with no history of neurological, cardiovascular, or psychiatric disorders. After excluding one participant due to excessive head motion during scanning (mean framewise displacement > 0.5 mm), 99 participants were retained for analysis (mean age = 66.38 ± 5.16 years; 17 males; mean years of education = 12.36 ± 2.85). This dataset was used to investigate associations between the principal functional connectivity gradient of the LPFC and executive function. The study was approved by the Ethics Committee of the Institute of Psychology, Chinese Academy of Sciences, and written informed consent was obtained from all participants before the study.

The lateral prefrontal cortex mask

According to Brodmann’s area classification, the LPFC includes BA 6, BA 8, BA 9, BA 10, BA 12/47, BA 44, BA 45, BA 46, and BA 9/46 [35]. These areas are divided into six main regions: premotor cortex, dorsolateral prefrontal cortex (DLPFC), inferior frontal cortex, frontopolar cortex, ventrolateral prefrontal cortex (VLPFC), and frontal eye fields [36]. Based on Brodmann labels detailed in the Human Brainnetome Atlas [37], we delineated a mask for the LPFC (Fig. 1A).

MRI acquisition

MRI data for the adult lifespan sample were obtained from the SALD. All scans were performed on a 3 T Siemens Trio scanner. High-resolution T1-weighted structural images were acquired using an MPRAGE sequence with the following parameters: TR = 1900 ms, TE = 2.52 ms, TI = 900 ms, flip angle = 90°, matrix size = 256 × 256, slices = 176, slice thickness = 1 mm, and voxel size = 1 × 1 × 1 mm³. Rs-fMRI data were obtained using a gradient-echo EPI sequence with: TR = 2000 ms, TE = 30 ms, flip angle = 90°, FOV = 220 × 220 mm², slices = 32, slice thickness = 3 mm with a 1 mm gap, and voxel size = 3.4 × 3.4 × 4 mm³. More detailed imaging parameters are available in previous literature [34].

To investigate the association between LPFC functional gradients and executive function, an independent cohort of older adults was scanned using a 3 T GE MR750 scanner. High-resolution T1-weighted structural images were acquired with the following parameters: TE = 2.928 ms, TI = 450 ms, flip angle = 12°, FOV = 256 × 256 mm², slice thickness = 1 mm, and voxel size = 1 × 1 × 1 mm³. Rs-functional images were obtained using a gradient-echo EPI sequence with: TR = 2000 ms, TE = 30 ms, Flip Angle = 90°, FOV = 224 × 224 mm², slices = 37, slice thickness = 3.5 mm, and voxel size = 3.5 × 3.5 × 3.5 mm³. Foam padding was applied to restrict head motion during scanning.

MRI data preprocessing

Rs-fMRI data were preprocessed using the Connectome Computation System (CCS) [38, 39]. The preprocessing followed a standardized sequence of operations: (1) removing the first 5 time points; (2) head motion correction; (3) slice timing correction; (4) outlier correction (de-spiking); (5) estimation of head motion parameters; (6) registering the functional images with the high-resolution T1 images using boundary-based registration; (7) removing nuisance factors such as head motion, cerebrospinal fluid, and white matter signals using ICA-based automatic removal of motion artifacts (ICA-AROMA); (8) removing linear and quadratic trends from the time series; (9) projecting the volumetric time series to surface space (the fsaverage5 model with medial wall masked out); (10) low-frequency band-pass filtering (0.01–0.1 Hz); and (11) 6 mm spatial smoothing. All preprocessing scripts are publicly accessible on GitHub (https://www.github.com/zuoxinian/CCS).

Connectome gradient analysis

The calculation of LPFC functional connectivity gradients was based on its resting-state functional connectivity to the entire cerebrum. According to a previous study [40], the vertex-level LPFC-to-cerebrum functional connectivity matrix (3287 × 20,484) was generated by calculating the Pearson correlation coefficients between each vertex within the LPFC and the time course of each vertex within the cerebrum, based on the individual-level BOLD time course. This matrix was then Fisher Z-transformed and thresholded to retain only the top 10% of connections per row [13, 20, 24, 41]. A cosine similarity matrix was then calculated to capture the similarity of connectivity profiles between each pair of LPFC vertices. We applied diffusion map embedding [13, 15] to calculate LPFC functional connectivity gradients. Diffusion map embedding is a nonlinear dimensionality reduction technique that estimates low-dimensional embeddings from high-dimensional connectome matrices. In the embedding space, cortical vertices that are densely connected (either by many connections or a few very strong connections) are positioned close to each other, whereas vertices with few or no interconnections are positioned farther apart. This approach belongs to the family of graph Laplacians and is derived from the equivalence between the Euclidean distance in the embedding space and the diffusion distance between probability distributions centered on these points. The algorithm is controlled by a single parameter α, which determines the influence of the density of sampling points on the underlying manifold (α = 0 for maximal effect; α = 1 for no effect). Based on prior studies [13, 20, 24, 41], we set α = 0.5 to preserve the global relationship between data points in the embedding space. First, we calculated the average functional connection matrix for each group (young, middle-aged, and older adults). Cortical gradient patterns for each age group were derived using diffusion map embedding applied to their respective connectivity matrices, and these were then mapped onto the cortical surface to visualize macroscale transitions in the overall connectivity pattern. To ensure that gradient axes were comparable across individuals, a group-level gradient component template was generated from the average connectivity matrix across all participants. Individual gradient components were spatially aligned to this template using Procrustes rotation [42], a process iterated 100 times [17]. Gradient components were sorted in descending order based on the proportion of variance they explained across the connectivity profiles. The principal gradient, which captured the greatest proportion of variance, was selected for subsequent statistical analyses.

Global gradient metrics comparison across the young, middle-aged, and older adults

We analyzed three global gradient metrics—explanation ratio, gradient range, and gradient variation—using a one-way ANOVA to elucidate their statistical characteristics. Specifically, the explanation ratio quantifies the proportion of the total variance accounted for by each gradient and is calculated by dividing the eigenvalue of a particular gradient by the sum of all eigenvalues [17]. This provides a numerical metric for the relative significance of each gradient within the overall data structure. The gradient range is measured as the difference between the maximum and minimum values observed within each gradient [43]. This metric is crucial for understanding the spatial extent of variation represented by the gradient across the dataset. Gradient variation, on the other hand, is defined as the standard deviation of the gradient values [43]. Both gradient variation and range were computed across all vertices within the LPFC mask, representing the within-region variability and range of gradient values. Although both gradient range and gradient variation describe the spatial distribution of gradient values, they capture distinct statistical properties. The range is defined solely by the two extreme values, whereas the variation incorporates information from all values across the region. Consequently, two gradients can possess similar ranges but markedly different variations, and vice versa. These metrics have also been adopted in prior studies applying gradient-based analyses across diverse populations—for example, in developmental research [43], individual-difference studies on creativity [17], and clinical investigations of depression [44]. Both gradient variation and range were computed across vertices within the LPFC. These works collectively support the interpretive validity and methodological reliability of using these metrics.

Subdivisions-level gradient comparison across the young, middle-aged, and older groups

To further investigate LPFC hierarchical changes in older adults, we compared the gradient values of specific subdivisions across young, middle-aged, and older adults. The average gradient value of each region was calculated for each participant, and group differences were assessed using a one-way ANOVA. The results were corrected for multiple comparisons using the false discovery rate (FDR) correction method.

Relative distance comparison across the young, middle-aged, and older adults

We assessed the relative distances between LPFC subdivisions along the principal gradient across age groups. Following the previous study [17], relative distance was defined as the absolute rank difference in the mean gradient values between each pair of LPFC subdivisions. Specifically, empirical gradient values were first transformed into rank-based gradient values within each participant, and then the absolute difference between the ranked values of two regions was computed. This measure reflects the degree of functional differentiation between regions: lower distances indicate greater similarity or functional proximity within the hierarchical axis, whereas higher distances reflect greater segregation or distinct hierarchical positions.

Meta-analytic functional decoding using NeuroSynth

To examine the cognitive function associated with age-related changes in LPFC connectivity gradients, we conducted a meta-analytic decoding analysis using the NeuroSynth database (www.neurosynth.org), which aggregates activation data from thousands of published fMRI studies. This approach aimed to identify cognitive terms most strongly related to the brain regions showing significant gradient alterations in older adults. We first generated two additional masks based on the regions that showed significant results in the subdivision-level comparative analysis. These masks were registered to MNI152 standard space at a 2 mm isotropic resolution and submitted to the NeuroSynth decoder. For each mask, we extracted the average z-statistics corresponding to associated cognitive feature terms, thus linking altered gradient patterns to meta-analytically derived cognitive functions.

Linking gradient alterations to executive function

To examine the behavioral relevance of LPFC functional gradients, we specifically analyzed the relationship between these gradients and executive function performance using this independently collected sample of 99 healthy older adults. Executive function was measured using the FISHERMAN system, a validated cognitive battery developed by our team to assess three core domains: inhibition, shifting, and working memory [45]. Raw scores from each domain were standardized into Z-scores and averaged to derive a composite executive function score for each participant. To assess the association between LPFC gradients and executive function, we computed Spearman correlation coefficients between the composite executive function score and three global gradient metrics: explanation ratio, gradient range, and gradient variation. To explore regional specificity, we also calculated the mean gradient value for each LPFC subdivision and correlated these values with participants’ executive function scores. All correlation analyses were controlled for age, gender, years of education, and head motion as covariates, and multiple comparisons were corrected via the FDR correction method with a significance threshold of p < 0.05. To further characterize these relationships, we conducted independent-sample t-tests comparing global and regional gradient metrics between participants in the highest and lowest quartiles (top and bottom 27%) of executive function performance (n = 27 per group). Age differed significantly between the groups (higher executive function group: 64.89 ± 4.75 years, lower executive function group: 68.93 ± 5.67 years, p < 0.05), while gender distribution did not reach significance (23 vs. 24 females in the higher and lower executive function groups, respectively, p > 0.05). Years of education (higher executive function group: 13.04 ± 2.85 years, lower executive function group: 11.59 ± 3.10 years, p = 0.081) and head motion, as indexed by mean framewise displacement (higher executive function group: 0.13 ± 0.06 mm, lower executive function group: 0.17 ± 0.09 mm, p = 0.065), showed marginally significant differences. In light of these findings, we included age, years of education, and head motion as covariates in all subsequent statistical analyses. This allowed us to identify gradient features that distinguish higher- vs. lower-performing individuals, offering insights into how LPFC hierarchical organization supports cognitive aging.

Results

Principal gradient patterns of the LPFC in young, middle-aged, and older adults

The principal gradient explained 39.28% of the variance in the LPFC-to-cerebrum connectome (young adults, 38.00%; middle-aged adults, 41.15%; older adults, 38.71%) (Fig. 1C). The principal gradient was anchored at one end by caudal LPFC regions and at the other end by rostral LPFC regions, supporting a dominant rostro-caudal organization in the LPFC (Fig. 2A). The rostro-caudal spatial layout of the principal gradient was clearly expressed by the posterior-anterior positions of the six brain regions (Fig. 2B). The spatial patterns of the group-averaged principal gradient maps were remarkably similar among young, middle-aged, and older adults (Spearman’s rs = 0.999, ps < 0.0001, permutation tests corrected for spatial autocorrelation, Fig. 2C).

Fig. 2.

Connectome gradient mapping across age groups. A The principal gradient was organized along a rostro-caudal axis within the lateral prefrontal cortex (LPFC). B Principal gradient distribution across LPFC subdivisions: premotor cortex (PMC), inferior frontal cortex (IFC), frontal eye fields (FEF), ventrolateral prefrontal cortex (VLPFC), dorsolateral prefrontal cortex (DLPFC), and frontopolar cortex (FPC). C Vertex-wise spatial correlations of group-averaged principal gradient maps: young vs. middle-aged (left), young vs. older adults (middle), middle-aged vs. older adults (right). High spatial similarity was observed across age groups

Comparison of global gradient metrics among young, middle-aged, and older adults

A one-way ANOVA indicated a significant group difference in the explanation ratio (F = 7.50, p < 0.01, partial η² = 0.031) and the gradient variation (F = 3.51, p < 0.05, partial η² = 0.015). Post-hoc analysis revealed that middle-aged adults had a significantly higher explanation ratio (p < 0.001) and greater gradient variation (p = 0.030) than young adults (Fig. 3A). However, no significant differences were observed between older and young adults, nor between older and middle-aged adults (ps > 0.05). There were no significant differences in gradient range (F = 1.74, p = 0.178, partial η² = 0.007) among the three groups.

Fig. 3.

A Group differences in global gradient metrics and relative distances (FDR correction). Middle-aged adults show higher gradient explained ratio and variation than young adults. Young adults exhibit a greater PMC-FPC distance compared to older adults. B Subdivisions-level gradients across age groups (FDR correction). Older adults showed significantly lower gradient values in the dorsolateral prefrontal cortex (DLPFC) and frontopolar cortex (FPC) compared to middle-aged adults. C NeuroSynth decoding of cognitive functions associated with altered principal gradient values in older adults

Comparison of subdivisions-level gradients among young, middle-aged, and older adults

In the principal gradient analysis, a one-way ANOVA revealed significant differences among the three groups in the DLPFC (F = 4.18, p < 0.05, partial η² = 0.017), frontopolar cortex (F = 5.30, p < 0.05, partial η² = 0.022). Post-hoc tests showed that older adults exhibited significantly lower gradient values in the DLPFC (p = 0.035) and frontopolar cortex (p = 0.004) compared to middle-aged adults (Fig. 3B). No significant differences were observed in other group comparisons of LPFC gradient values.

Comparison of relative distance among young, middle-aged, and older adults

A one-way ANOVA revealed significant differences in the relative distances between the premotor cortex and frontopolar cortex (F = 3.97, p < 0.05, partial η² = 0.016). Post-hoc comparisons showed that the relative distances between the premotor cortex and frontopolar cortex were significantly higher in young adults than in older adults (p < 0.05). Furthermore, a marginally significant difference was observed between middle-aged and older adults (p = 0.054) (Fig. 3A).

NeuroSynth decoding of gradient-altered LPFC subdivisions

To explore the functional relevance of age-related changes in LPFC connectivity gradients, we performed a meta-analytic decoding analysis using NeuroSynth. Two masks were created based on subdivision-level comparisons: DLPFC and frontopolar cortex. The NeuroSynth analysis revealed that DLPFC and frontopolar cortex were primarily associated with memory terms, such as working memory, episodic memory, recognition memory, and autobiographical memory (Fig. 3C).

Correlations between gradient metrics and executive function performance in older adults

We next examined how LPFC gradient metrics relate to executive function in older adults. At the global level, all three global gradient metrics—explanation ratio, gradient range, and gradient variation—were significantly positively correlated with core components of executive function, including inhibition (r = 0.30–0.33, p < 0.01, r² = 0.089–0.100), shifting (r = 0.38, p < 0.01, r² = 0.141–0.143), and the composite executive function Z-score (r = 0.36–0.37, p < 0.01, r² = 0.130–0.136) (Figs. 4A–C). All these correlations remained significant after FDR correction.

Fig. 4.

LPFC gradient-executive function correlations in older adults. A–C Positive correlations between executive function and global gradient metrics (explanation ratio, range, variation). D–F Correlations with regional gradient values, showing D a negative association for the PMC and E, F positive associations for the DLPFC and VLPFC. Scatterplots show FDR-corrected Spearman’s partial correlations (controlling for age, gender, education, and head motion), visualizing residual variance

At the subdivision-level, specific LPFC regions showed differential associations with executive performance. Gradient values in the premotor cortex (PMC) were negatively correlated with inhibition (r = −0.32, p < 0.01, r² = 0.104), shifting (r = −0.28, p < 0.05, r² = 0.076), and the composite executive function score (r = −0.32, p < 0.01, r² = 0.103) (Fig. 4D). In contrast, positive correlations were observed in the DLPFC with inhibition (r = 0.35, p < 0.01, r² = 0.126), shifting (r = 0.32, p < 0.01, r² = 0.103), and the composite score (r = 0.33, p < 0.01, r² = 0.106) (Fig. 4E), reflecting their continued role in supporting executive control. Additionally, gradient values in the VLPFC area were positively associated with the composite score (r = 0.25, p < 0.05, r² = 0.064) (Fig. 4F). All these correlations remained significant after FDR correction.

To further probe these relationships, we conducted group comparisons between individuals in the top and bottom 27% of the executive function score distribution (n = 27 per group). Participants in the higher executive function group exhibited significantly higher global gradient metrics, including explanation ratio (t = 4.52, p < 0.001), gradient range (t = 4.52, p < 0.001), and gradient variation (t = 4.22, p < 0.001) (Fig. 5A). At the subdivision level, the higher executive function group also showed higher gradient values in the DLPFC (t = 3.85, p < 0.01) and VLPFC (t = 2.83, p < 0.05) and lower values in the PMC (t = –3.97, p < 0.01) compared to their lower-performing counterparts (Fig. 5B).

Fig. 5.

Executive function subgroup comparisons of LPFC gradient metrics. A Higher executive function subgroup showed significantly greater explanation ratio, gradient range, and variation than the lower subgroup. B At the regional level, the higher subgroup exhibited higher gradient values in DLPFC/VLPFC but lower values in PMC versus lower subgroup. All comparisons were FDR corrected

Discussion

Gradient analysis provides a comprehensive framework for characterizing the hierarchical functional organization of the human brain. When applied to specific cortical territories, this approach reveals the spatial organization and internal differentiation of connectivity patterns within the region [13, 46]. In the present study, we investigated age-related differences in the primary gradient of functional connectivity within the LPFC. Our results revealed that, although the overall rostro-caudal organization of LPFC gradients remains largely stable across the adult lifespan, specific localized age-related alterations are evident. In particular, gradient values in the rostral LPFC, including regions such as the frontopolar cortex, were significantly reduced in older adults relative to middle-aged individuals. Moreover, the relative distance between anterior high-gradient regions (e.g., frontopolar cortex) and posterior low-gradient regions (e.g., premotor cortex) was significantly diminished in older adults, indicating a compression of the LPFC’s functional connectivity gradient with age.

These findings support the notion that the LPFC retains a stable hierarchical structure—from posterior regions involved in sensory and motor processing to anterior regions supporting abstract, goal-directed cognition—throughout adulthood. This preserved rostro-caudal gradient mirrors prior observations of age-invariant functional organization at the whole-brain level [24, 25], suggesting that core hierarchical principles of cortical architecture may be resilient to age-related degeneration. Despite known patterns of structural atrophy and functional reorganization in aging, the persistence of this gradient structure implies a possible compensatory mechanism aimed at maintaining cognitive function. These results align with broader research showing that, while localized connectivity alterations do occur, the fundamental macroscale gradient architecture of the brain remains intact across the lifespan [24, 25, 47].

Our findings extend beyond whole-brain gradient descriptions by revealing distinct LPFC-specific reorganization dynamics. While middle-aged adults exhibited a higher explanation ratio and greater gradient variability than young adults—potentially reflecting adaptive reorganization—older adults showed significant compression in dorsolateral and frontopolar subregions relative to their middle-aged counterparts. This non-monotonic pattern, characterized by preserved young-old similarity but prominent midlife-to-old-age differences, suggests that LPFC gradient alterations accelerate during the transition from mid-to-late adulthood rather than following a linear progression. This regional specificity provides important insights beyond whole-brain models: the LPFC appears to undergo particularly dynamic changes during midlife, potentially serving as an early indicator of subsequent cognitive decline. The absence of significant young-old differences, coupled with pronounced midlife changes, reinforces the value of including middle-aged participants and supports a stage-dependent model of LPFC reorganization. These findings position midlife as a critical period for early detection and intervention, with LPFC gradient metrics offering regionally specific markers that complement whole-brain aging models. It is important to note that functional connectivity gradients do not refer to the strength of functional connections. Rather, they represent low-dimensional representations of whole-brain resting-state connectivity patterns between specific regions and the entire cortex [16]. The DLPFC and the frontopolar cortex occupy the most anterior positions along the LPFC’s functional hierarchy, and their lower gradient values reflect progressive spatial convergence along the principal gradient axis. The observed anterior–posterior compression pattern may reflect a weakened hierarchical control gradient within the LPFC. Typically, anterior regions integrate information across temporal and representational scales to support abstract, goal-directed behavior [48, 49]. This integrative capacity appears to deteriorate with aging, leading to reduced differentiation between anterior and posterior control processes. This disruption may represent a loss of dynamic flexibility in coordinating concrete sensorimotor and abstract control systems, consistent with the neural dedifferentiation hypothesis of aging [50]. Importantly, our findings extend beyond prior whole-brain or meta-analytic studies by demonstrating that LPFC-specific gradient compression is uniquely associated with variability in executive performance, highlighting the potential of LPFC functional gradients as a regionally specific biomarker of cognitive aging.

At the whole-brain level, previous studies have demonstrated compression of the principal gradient extending from primary sensorimotor areas to the default mode network in older adults compared to younger individuals [25, 26]. This supports the view that age-related alterations in functional network architecture are characterized by decreased within-network connectivity and increased between-network integration [51–53]. Specifically, older adults exhibit enhanced functional connectivity between frontoparietal control systems and other neural networks, reflecting reduced functional specialization with advancing age [51]. In parallel with these global patterns, our findings of LPFC gradient compression and lower spatial separation between the premotor cortex and the frontopolar cortex in older adults may reflect adaptive neurocognitive mechanisms. The reduced functional distance between networks may reflect compensatory mechanisms that enhance cognitive resource mobilization, thereby helping to maintain behavioral performance despite age-related structural decline.

The principal gradient of cortical functional connectivity has been shown to effectively capture a continuum of cognitive processes, ranging from basic perceptual functions to high-level abstract cognition [13]. In the present study, we found that higher explanation ratios, greater gradient ranges, and higher gradient variability were all associated with better executive function. Specifically, individuals with better executive function exhibited higher gradient values in the DLPFC and VLPFC, and lower gradient values in the PMC. Given that distance along the principal gradient reflects the similarity or dissimilarity of a region’s connectivity profile relative to the rest of the brain, greater gradient range and variability suggest more pronounced functional differentiation within the LPFC. This differentiation appears to be beneficial for sustaining executive function in aging. These findings are consistent with the tethering hypothesis, which posits that transmodal association cortices develop their integrative and abstract processing capabilities by increasing their functional distance from primary sensory regions, thereby reducing constraint from external sensory inputs [16, 54]. The rostro-caudal gradient within the LPFC, which has been proposed as a scaffold for executive control [49], reflects this hierarchy: posterior regions such as the PMC are primarily involved in sensorimotor processing, while mid-LPFC areas handle contextual integration, and anterior regions (e.g., DLPFC and frontopolar cortex) support abstract reasoning and goal representation. A greater anterior–posterior gradient distance may thus functionally decouple higher-order cognitive systems from lower-level sensory inputs, facilitating strategic thinking, abstract reasoning, and top-down control—all of which are essential for effective executive functioning, particularly in late adulthood.

By integrating connectome gradient analysis with lifespan developmental neuroscience, this study advances our understanding of the neural mechanisms underlying cognitive aging. The observed gradient compression in the LPFC supports the dedifferentiation hypothesis, which posits that aging leads to reduced functional segregation and specialization across brain networks [55, 56]. This functional dedifferentiation is thought to be driven by cumulative structural, vascular, metabolic, and neurophysiological changes associated with aging [57, 58], and aligns with prior evidence of less distinct activation patterns in aging brains [59–61]. Beyond the structural and functional changes discussed, emerging evidence suggests that vascular insufficiency and neuroinflammation may represent key drivers of gradient compression. Recent findings indicate that older adults with more compressed gradients exhibit poorer vascular health, and that vascular features modulate the relationship between gradient values and cognitive performance [25]. This suggests that compromised neurovascular coupling and chronic low-grade inflammation may synergistically accelerate functional dedifferentiation. Future studies integrating multimodal imaging—including arterial spin labeling for cerebral blood flow, diffusion tensor imaging for white matter integrity, and positron emission tomography for neuroinflammation—will be crucial to elucidate the multifactorial mechanisms underlying gradient reorganization. While LPFC gradient compression shows promise as a biomarker of cognitive aging, several validation steps are needed before clinical implementation. Future work should establish its predictive validity for longitudinal cognitive decline through prospective cohort studies, test its sensitivity to interventions (e.g., cognitive training and physical activity) in randomized controlled trials, and assess its feasibility for integration into standardized clinical neuroimaging protocols by determining scan-rescan reliability and developing automated analysis pipelines. These steps will determine whether gradient metrics can ultimately serve as clinically useful tools for early risk detection and intervention monitoring.

Several limitations of the present study warrant consideration. First, although the cross-sectional design reveals age-dependent variations in the LPFC functional hierarchy, this cross-sectional nature of our study fundamentally limits causal interpretation of the observed effects. While revealing clear age-related differences, this design cannot definitively disentangle neurobiological aging from potential cohort effects stemming from generational or lifestyle differences. Such confounding may substantially influence the observed gradient compression patterns. Future longitudinal studies tracking within-individual changes across time will be essential to determine whether these patterns reflect genuine aging processes or cohort-specific influences. Second, while the NeuroSynth meta-analytic decoding framework offers a valuable means of linking brain activation patterns to cognitive functions, it yields probabilistic and indirect associations. The inferred cognitive relevance of gradient alterations may therefore lack domain specificity, and future work should corroborate these findings using task-based fMRI and behavioral validation. Third, our investigation of the relationship between LPFC gradients and executive function was limited to older adults, restricting the generalizability of the findings across the broader adult lifespan. Future work should explore whether similar gradient–cognition associations are present in young and middle-aged populations, thereby elucidating lifespan trajectories of functional reorganization and cognitive resilience. Finally, the exclusive reliance on resting-state fMRI data constrains both the interpretative scope and ecological validity of the findings. Incorporating multimodal neuroimaging approaches—such as combining resting-state and task-based fMRI with diffusion-weighted imaging (DWI)—would enable a more comprehensive investigation of the interrelationships among functional hierarchy, structural connectivity, and cognitive performance. Such integrative methods would not only enhance the neurobiological interpretability of gradient metrics but also improve the clinical translatability of findings by capturing brain function under more behaviorally relevant conditions.

Conclusion

This study characterizes age-related changes in the hierarchical functional architecture of the LPFC using connectome gradient mapping. While the rostro-caudal gradient organization remains largely stable across adulthood, older adults exhibit significant gradient axis compression, particularly in the DLPFC and frontopolar cortex. This compression, marked by reduced functional differentiation and diminished spatial separation, may underlie cognitive decline. Furthermore, observed gradient alterations correlate with cognitive functions such as observation, execution, and emotion regulation, suggesting a mechanistic link between LPFC reorganization and aging-related cognitive decline.

Author contribution

HYH: conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, and writing—original draft; PW: data curation, investigation, and methodology. SJG: data curation and investigation; HJL: conceptualization, supervision, and writing—review and editing.

Funding

This study is supported by the National Natural Science Foundation of China (32471121, 32500945), the National Basic Science Data Center “Chinese Data-sharing Warehouse for In-vivo Imaging Brain” (NBSDC-DB-15), and the project for enhancing young and middle-aged teacher’s research basis ability in colleges of Guangxi (2025KY0439). The funders played no role in the design, conduct, or reporting of this study.

Software and code availability

All image preprocessing scripts are publicly accessible on GitHub (https://www.github.com/zuoxinian/CCS). Gradient analysis and meta-analytic decoding were performed in MATLAB R2022b and Python 3.6.13, respectively, following the framework of Margulies et al. (13) (https://github.com/NeuroanatomyAndConnectivity/gradient_analysis). Group-level statistical analyses and correlation analyses were performed in R version 4.3.2. The Southwest University Adult Lifespan Dataset (SALD) used in this study is publicly available (https://doi.org/10.15387/fcp_indi.sald). The independently collected imaging and behavioral data of older adults can be obtained by contacting the corresponding author.

Declarations

Ethics approval and consent to participate

Human subjects: The experiments were approved by the Institutional Review Board of the Institute of Psychology, Chinese Academy of Science . Informed consent was obtained from all subjects. SALD dataset: This is open data; the data collection was approved by the Research Ethics Committee of the Brain Imaging Center, Southwest University, in compliance with the Declaration of Helsinki.Written informed consent was obtained from all SALD participants at the time of original data acquisition (Wei et al., Sci Data. 2018, reference 34).

Conflict of interest

The authors declare no competing interests.