HABS-HD: Health Disparities in Cognitive Aging

• HABS-HD is a large, multiethnic cohort study that integrates imaging, genetic, and clinical data to examine Alzheimer’s and cognitive aging disparities.
• It employs advanced causal modeling with ICA-LiNGAM and directed hyperconnectome frameworks to uncover shared and population-specific neural circuit dynamics.
• Findings underscore that subgroup analyses and predictive modeling enhance precision interventions tailored to neurobiological and psychosocial contexts in aging.

The Health and Aging Brain Study Health Disparities (HABS-HD) comprises a large-scale, multiethnic cohort designed to address critical gaps in Alzheimer’s Disease (AD) and cognitive aging research arising from the underrepresentation of minority populations in neuroimaging studies. HABS-HD integrates population, imaging, genetic, and clinical data to enable causal and subgroup analyses of brain network architecture, disease progression, psychological risk factors, and treatment effects, with explicit emphasis on health disparities.

1. Cohort Composition, Demographics, and Modalities

HABS-HD enrolled 3,840 participants (mean age 64.82 ± 8.19 years; 2,379 female) systematically sampled across three self-identified groups: African American (AA, N=1,066), Hispanic (N=1,425), and Non-Hispanic White (NHW, N=1,349). Participants span clinical disease stages, including Normal Control (NC), Mild Cognitive Impairment (MCI), and Alzheimer’s Disease (AD), classified by multidisciplinary consensus. In addition to AD staging, the Penn State Worry Questionnaire (PSWQ) provides a dimensional measure of psychological risk (trait worry), dichotomized at PSWQ ≥ 60 (high worry) versus < 60 (low worry).

Functional neuroimaging is conducted via resting-state fMRI (rs-fMRI), with a single session (150 time points) per subject. Standardized preprocessing is performed using the CONN toolbox, including realignment, normalization to Montreal Neurological Institute (MNI) space, spatial smoothing, nuisance regression for motion and white matter/cerebrospinal fluid, and band-pass filtering (0.008–0.09 Hz). The brain is parcellated into N = 132 regions of interest (ROIs), combining Harvard-Oxford and AAL atlases. Imaging data are processed identically across sites, without explicit harmonization, to minimize batch effects (Moreno et al., 2 Nov 2025).

2. Causal Modeling and Directed Hyperconnectome Framework

HABS-HD analyses leverage a subject-specific causal model at the ROI level, implemented via Independent Component Analysis–Linear Non-Gaussian Acyclic Model (ICA-LiNGAM). For each time series $x_t \in \mathbb{R}^{N}$ (BOLD signal at time $t$):

$(1)\quad x_t = B x_t + e_t$

where $B \in \mathbb{R}^{N\times N}$ is a strictly lower-triangular causal adjacency matrix, and $e_t$ is non-Gaussian noise. ICA yields $X = M S$ (mixing and source matrices), $W = M^{-1}$ (unmixing), and the causal structure is recovered by:

$(2)\quad B = I - W^{-1}$

To link ROI-level interactions to system-level organization, ROIs are mapped onto $K=11$ pre-defined functional systems (Yeo’s 7 networks plus cerebellum, vermis, brainstem, and subcortical). For systems $S_k$ and $S_\ell$, the directed hyperconnectome matrix $H$ is defined:

$$(3)\quad H_{k\ell} = \frac{1}{|S_k||S_\ell|} \sum_{i\in S_k} \sum_{j\in S_\ell} B_{ij}$$

This approach yields third-order tensors $H^{(s)}_{k\ell}$ (subjects × systems × systems), supporting within- and between-population comparative analyses (Moreno et al., 2 Nov 2025).

3. Population-Specific Disease and Risk Mechanisms

Within HABS-HD, population-aware analyses partition individuals according to self-identified race/ethnicity and derive group-specific hyperconnectome patterns. Statistical significance of inter-group differences in directed edges (system-wise) is assessed via permutation testing ($n$ = 5,000 label randomizations; Bonferroni correction for 110 tests).

Population-specific findings include the identification of both shared and unique directed connectivity patterns associated with AD progression and trait worry. For example, four MCI-related hyperedges—FPN→VIS, FPN→DMN, Brainstem→DMN, Brainstem→Cerebellum—are robust across all groups with effect sizes $\Delta \approx 0.05\text{–}0.08$ and $p < 10^{-4}$. However, population-dependent variations are observed: AA AD cases present uniquely weakened DAN→Cerebellum connectivity ($p=0.0027$) and stronger limbic→FPN influences; in Hispanic individuals, the directionality of Brainstem–VAN/SAL edges reverses between MCI and high-worry subgroups ($\Delta_\text{dir}\approx0.06$, $p<0.001$), suggestive of shifting arousal–salience network modulation (Moreno et al., 2 Nov 2025).

4. Directed Closed-Loop Circuits and Network Regulation

Analysis of the directed hypergraph structure reveals simple cycles (length $m\geq3$) in the system-level connectome, interpreted as closed-loop regulatory circuits. Detection uses depth-first search on the adjacency $H$; statistical significance is evaluated against null distributions from permuted labels.

Group-specific loops include: (i) a lower-level arousal–motor loop in Hispanic AD (Brainstem→SMN→Subcortical→Brainstem), potentially reflecting increased reliance on sensorimotor and subcortical regulatory circuits as cortical function deteriorates; (ii) a somatic–rumination loop in AA high-worry cases (SMN→DMN→SMN), aligning with hypotheses that bodily tension potentiates perseverative self-referential states. Such circuits are not uniformly present across all groups or phenotypes, underscoring population-dependent regulatory strategies in both cognitive and affective domains (Moreno et al., 2 Nov 2025).

5. Predictive Modeling and Population-Aware Implications

Classification accuracy for early AD progression (NC/MCI/AD) and trait worry (high/low PSWQ) is enhanced by modeling population-specific hyperconnectome variation. Multilayer perceptron (MLP) performance on disease classification achieves pooled accuracy $72.03\pm1.59\%$ (NHW $85.59\pm1.91\%$, Hispanic $73.10\pm2.04\%$, AA $70.52\pm1.89\%$). Trait worry classification reaches $84.70\pm1.87\%$ pooled accuracy (NHW $89.03\pm1.36\%$, Hispanic $88.43\pm3.06\%$, AA $87.60\pm2.44\%$). Neglecting race-specific differences yields systematic performance loss, indicating the necessity of population-aware neuroimaging models. Common vulnerability pathways—particularly FPN→DMN and brainstem-driven influences—are consistently implicated, but regulatory circuits and effect directions can be population- and context-dependent (Moreno et al., 2 Nov 2025).

6. Subgroup Analysis, Treatments, and Health Disparities

Advanced causal subgroup discovery methods, such as fused optimal causal trees (FOCT), have been specifically applied to HABS-HD to identify heterogeneous treatment effects among Black older adults (N=237). The FOCT framework partitions the covariate space—modeled as:

$$Y_i = \sum_{m=1}^{M(\Pi)}\mathbf{1}_{\{X_i\in\mathcal{A}_m\}} \left(\delta_m+\mu_m T_i+\alpha_m^\top X_i+\beta_m^\top (T_i X_i)\right) + \epsilon_i$$

where $Y$ is standardized MMSE, $T$ is binary Aβ-treatment indicator (Pittsburgh compound-B PET centiloid $\geq 10$ vs. $<10$), and $\boldsymbol{X}$ contains sex, APOE alleles, education, age, entorhinal and neocortical tau PET. The optimization problem includes a global parameter fusion constraint, enabling information sharing across related subgroups and controlling statistical efficiency via an $\ell_0$ fusion penalty.

In the HABS-HD analytic sample, four subgroups were identified based on splits in neocortical tau ($\mathrm{Tau}_2$), education, and age. Only the “high-tau, low-education” group ($\mathrm{Tau}_2 \geq 1.071$, Edu < 15) demonstrated a statistically significant cognitive improvement from Aβ-lowering treatment ($\hat{\beta}_{\mathrm{Tau}_2}=-0.962$, $p<0.001$). Fusion constraints resulted in narrower confidence intervals (30–40% shorter) and stabilized estimates, especially in underrepresented subgroups (Xie et al., 3 Feb 2026).

FOCT findings suggest that individuals with high neocortical tau and low education represent a priority subgroup for targeted intervention and that education exerts a strong protective effect, supporting cognitive-reserve models. Recommendations include pre-specifying tau/education subgroups in trial designs and extending parameter fusion strategies to improve power and accuracy in health disparities research.

7. Implications for Precision Medicine and Future Directions

Analyses performed on HABS-HD demonstrate that both shared network vulnerabilities (notably FPN→DMN, brainstem → DMN/Cerebellum) and population-dependent regulatory differences underlie cognitive decline and risk phenotypes. The detection of group-specific closed-loop circuits and context-dependent edge reversals across disease and affective states point to adaptive differences shaped by lifelong disparity exposures and sociocultural factors.

A plausible implication is that interventions in AD must be tailored to the neurobiological and psychosocial contexts of diverse populations. The integration of causal modeling with health-disparity-aware subgroup analysis, as exemplified by HABS-HD, is essential for the development of precise interventions, enhanced external validity, and equitable translational pipelines in brain aging and dementia research (Moreno et al., 2 Nov 2025, Xie et al., 3 Feb 2026).