Statistical Advances in Neuroimaging and Cognitive Health

We explore methods to reduce the impact of unobserved confounders on the causal mediation analysis of high-dimensional mediators with spatially smooth structures, such as brain imaging data. The key approach is to incorporate the latent individual effects, which influence the structured mediators, as unobserved confounders in the outcome model, thereby potentially debiasing the mediation effects. We develop BAyesian Structured Mediation analysis with Unobserved confounders (BASMU) framework, and establish its model identifiability conditions. Theoretical analysis is conducted on the asymptotic bias of the Natural Indirect Effect (NIE) and the Natural Direct Effect (NDE) when the unobserved confounders are omitted in mediation analysis. For BASMU, we propose a two-stage estimation algorithm to mitigate the impact of these unobserved confounders on estimating the mediation effect. Extensive simulations demonstrate that BASMU substantially reduces the bias in various scenarios. We apply BASMU to the analysis of fMRI data in the Adolescent Brain Cognitive Development (ABCD) study, focusing on four brain regions previously reported to exhibit meaningful mediation effects. Compared with the existing image mediation analysis method, BASMU identifies two to four times more voxels that have significant mediation effects, with the NIE increased by 41%, and the NDE decreased by 26%. 

Scalp event-related potentials (ERPs) measured with electroencephalography (EEG) are temporally precise but spatially bandwidth-limited: skull and scalp blur cortical signals, and ERP topographies are typically described in electrode space rather than in coordinates tied to cortical anatomy. We introduce a cortex-anchored sensor-space basis by forward-projecting cortical Laplace-Beltrami (LB) eigenmodes through a realistic head model, yielding a fixed multiscale dictionary whose ordering inherits a cortical spatial-frequency structure. We compute LB eigenmodes on the fsaverage template, map them to sensors via a three-layer boundary-element head model, and define a reusable group-montage dictionary. We compare this forward-projected LB basis to (i) spherical harmonics (SPH) on the same montage and (ii) group PCA/ICA bases trained on trial-averaged time-frequency (TF) maps. The results support forward-mapped LB eigenmodes as a compact, anatomy-anchored coordinate system for resting and evoked EEG that complements spherical and data-adaptive bases and provides a general-purpose feature space for geometry-aligned EEG analyses. 

Psychiatric disorders affect more than 20% of the population in the United States. Advances in neuroscience research have shown the potential clinical uses of neuroimaging markers for diagnosis, treatment selection, and disease monitoring. However, current clinical practice depends solely on clinical assessments without using biomarkers. The main obstacles are low test-retest reliability (TRR), the absence of rigorous TRR methods, methodological inconsistencies, and inadequate replication. To address this issue, we propose developing the Generalized Imaging Intraclass Correlation Coefficients (GI2C2) reliability framework designed for high-dimensional connectome and neuroimaging data. Our goal is to create statistically rigorous, clinically relevant, and scalable tools to help researchers and clinicians identify reproducible connectome biomarkers for psychiatric illnesses. 

Accurately characterizing cognitive decline across multiple domains on a clinically meaningful time scale is critical for understanding Alzheimer's disease (AD) progression, yet difficult to achieve because the disease process spans decades. Consequently, most available datasets consist of fragmentary individual data, which captures only random segments of an individual's potential progression. This can lead to conflicting conclusions when analyzed using standard strategies. To overcome these data challenges, we proposed a novel statistical framework, the double anchoring events-based sigmoidal mixed model (DSMM), and applied it to harmonized data from 13 cohorts within the ADSP-PHC. The DSMM characterizes cognitive decline trajectories relative to time-to-incident AD onset; uniquely, it introduces the concept of "secondary anchoring events" to effectively align trajectories for participants without an observed incident AD diagnosis. This innovation integrates fragmentary data into a cohesive model, significantly reducing selection bias while ensuring the replicability, interpretability, and representativeness of the resulting trajectories. Our results revealed a robust temporal order where memory decline generally precedes language impairment, followed by executive function. This pattern persists across subgroups defined by APOE ε4 status, sex, and race/ethnicity, highlighting how advanced statistical modeling of harmonized multi-cohort data can successfully resolve discrepancies in disease progression research. 

A fundamental challenge in psychiatric research is the lack of a coherent, empirically grounded framework for psychiatric nosology. Even within a single diagnostic category such as autism spectrum disorder (ASD), individuals exhibit substantial heterogeneity in symptoms, developmental trajectories, and neurobiological profiles. Traditional categorical definitions often fail to capture the continuous nature of underlying neurodevelopmental variation, motivating ongoing shifts toward dimensional classification systems. However, conventional statistical approaches typically assume multivariate normality, which may obscure complex, non-normal latent structures that characterize ASD-related variation.
To address this issue, we propose a semi-supervised independent component analysis (ICA)–based projection pursuit framework to identify latent axes of psychopathology based on patterns of statistical non-normality, such as skewness, multimodality, or U-shaped distributions. By integrating clinical diagnostic labels with functional connectivity data, the method aims to guide the ICA-based projections toward components that are both statistically non-Gaussian (due to psychopathology heterogeneity) and diagnostic-informative. The framework will be evaluated through simulation studies and applied to the ABIDE dataset to assess its utility in ASD nosology.