publications

Individualized treatment regimes (ITRs) aim to improve clinical outcomes by assigning treatment based on patient-specific characteristics. However, existing methods often struggle with high-dimensional covariates, limiting accuracy, interpretability, and real-world applicability. We propose a novel sufficient dimension reduction approach that directly targets the contrast between potential outcomes and identifies a low-dimensional subspace of the covariates capturing treatment effect heterogeneity. This reduced representation enables more accurate estimation of optimal ITRs through outcome-weighted learning. To accommodate observational data, our method incorporates kernel-based covariate balancing, allowing treatment assignment to depend on the full covariate set and avoiding the restrictive assumption that the subspace sufficient for modeling heterogeneous treatment effects is also sufficient for confounding adjustment. We show that the proposed method achieves universal consistency, i.e., its risk converges to the Bayes risk, under mild regularity conditions. We demonstrate its finite sample performance through simulations and an analysis of intensive care unit sepsis patient data to determine who should receive transthoracic echocardiography.

Functional magnetic resonance imaging (fMRI) provides an indirect measurement of neuronal activity via hemodynamic responses that vary across brain regions and individuals. Ignoring this hemodynamic variability can bias downstream connectivity estimates. Furthermore, the hemodynamic parameters themselves may serve as important imaging biomarkers. Estimating spatially varying hemodynamics from resting-state fMRI (rsfMRI) is therefore an important but challenging blind inverse problem, since both the latent neural activity and the hemodynamic coupling are unknown. In this work, we propose a methodology for inferring hemodynamic coupling on the cortical surface from rsfMRI. Our approach avoids the highly unstable joint recovery of neural activity and hemodynamics by marginalizing out the latent neural signal and basing inference on the resulting marginal likelihood. To enable scalable, high-resolution estimation, we employ a deep neural network combined with conditional normalizing flows to accurately approximate this intractable marginal likelihood, while enforcing spatial coherence through priors defined on the cortical surface that admit sparse representations. The proposed approach is extensively validated using synthetic data and real fMRI datasets, demonstrating clear improvements over current methods for hemodynamic estimation and downstream connectivity analysis.

Non-invasive measurements of the human brain using magnetic resonance imaging (MRI) have significantly improved our understanding the brain’s network organization by enabling measurement of anatomical connections between brain regions (structural connectivity) and their coactivation (functional connectivity). Heritability analyses have established that genetics account for considerable intersubject variability in structural and functional connectivity. However, characterizing how genetics shape the relationship between structural and functional connectomes remains challenging, since this association is obscured by unique environmental exposures in observed data. To address this, we develop a regression analysis framework that enables characterization of the relationship between latent genetic contributions to structural and functional connectivity. Implementing the proposed framework requires estimating genetic covariance matrices in multivariate random effects models, which is computationally intractable for high-dimensional connectome data using existing methods. We introduce a constrained method-of-moments estimator that is several orders of magnitude faster than existing methods without sacrificing estimation accuracy. For the genetic regression analysis, we develop regularized estimation approaches, including ridge, lasso, and tensor regression. Applying our method to Human Connectome Project data, we find that functional connectivity is moderately predictable from structure at the genetic level (max R^2 = 0.34), though it is not directly predictable in the observed data (max R^2 = 0.03). This stark contrast suggests that unique environmental factors mask strong genetically-encoded structure-function relationships.

Transcriptome-wide association studies (TWASs) are widely used to prioritize genes for diseases. Current methods test gene-disease associations at the bulk tissue or cell-type-specific pseudobulk level, which do not account for the heterogeneity within cell types. We present TWiST, a statistical method for TWAS at cell-state resolution using single-cell expression quantitative trait locus (eQTL) data. Our method uses pseudotime to represent cell states and models the effect of gene expression on the trait as a continuous pseudotemporal curve. Therefore, it allows flexible hypothesis testing of global, dynamic, and nonlinear associations. Through simulation studies and real data analysis, we demonstrated that TWiST leads to significantly improved power compared to pseudobulk methods. Application to the OneK1K study identified hundreds of genes with dynamic effects on autoimmune diseases along the trajectory of immune cell differentiation. TWiST presents great promise to understand disease genetics using single-cell studies.

Motivated by the problem of inferring the graph structure of functional connectivity networks from multi-level functional magnetic resonance imaging data, we develop a valid inference framework for high-dimensional graphical models that accounts for group-level heterogeneity. We introduce a neighborhood-based method to learn the graph structure and reframe the problem as that of inferring fixed effect parameters in a doubly highdimensional linear mixed model. Specifically, we propose a LASSO-based estimator and a de-biased LASSO-based inference framework for the fixed effect parameters in the doubly high-dimensional linear mixed model, leveraging random matrix theory to deal with challenges induced by the identical fixed and random effect design matrices arising in our setting. Moreover, we introduce consistent estimators for the variance components to identify subject-specific edges in the inferred graph. To illustrate the generality of the proposed approach, we also adapt our method to account for serial correlation by learning heterogeneous graphs in the setting of a vector autoregressive model. We demonstrate the performance of the proposed framework using real data and benchmark simulation studies.

We introduce a novel framework for the classification of functional data supported on nonlinear, and possibly random, manifold domains. The motivating application is the identification of subjects with Alzheimer’s disease from their cortical surface geometry and associated cortical thickness map. The proposed model is based upon a reformulation of the classification problem as a regularized multivariate functional linear regression model. This allows us to adopt a direct approach to the estimation of the most discriminant direction while controlling for its complexity with appropriate differential regularization. Our approach does not require prior estimation of the covariance structure of the functional predictors, which is computationally prohibitive in our application setting. We provide a theoretical analysis of the out-of-sample prediction error of the proposed model and explore the finite sample performance in a simulation setting. We apply the proposed method to a pooled dataset from the Alzheimer’s Disease Neuroimaging Initiative and the Parkinson’s Progression Markers Initiative. Through this application, we identify discriminant directions that capture both cortical geometric and thickness predictive features of Alzheimer’s disease that are consistent with the existing neuroscience literature.

Functional Principal Component Analysis (FPCA) has become a widely used dimension reduction tool for functional data analysis. When additional covariates are available, existing FPCA models integrate them either in the mean function or in both the mean function and the covariance function. However, methods of the first kind are not suitable for data that display second-order variation, while those of the second kind are time-consuming and make it difficult to perform subsequent statistical analyses on the dimensionreduced representations. To tackle these issues, we introduce an eigen-adjusted FPCA model that integrates covariates in the covariance function only through its eigenvalues. In particular, different structures on the covariate-specific eigenvalues—corresponding to different practical problems—are discussed to illustrate the model’s flexibility as well as utility. To handle functional observations under different sampling schemes, we employ local linear smoothers to estimate the mean function and the pooled covariance function, and a weighted least square approach to estimate the covariate-specific eigenvalues. The convergence rates of the proposed estimators are further investigated under the different sampling schemes. In addition to simulation studies, the proposed model is applied to functional Magnetic Resonance Imaging scans, collected within the Human Connectome Project, for functional connectivity investigation. Supplementary materials for this article are available online.

We introduce a framework for the reconstruction and representation of functions in a setting where these objects cannot be directly observed, but only indirect and noisy measurements are available, namely an inverse problem setting. The proposed methodology can be applied either to the analysis of indirectly observed functional images or to the associated covariance operators, representing second-order information, and thus lying on a non-Euclidean space. To deal with the ill-posedness of the inverse problem, we exploit the spatial structure of the sample data by introducing a flexible regularizing term embedded in the model. Thanks to its efficiency, the proposed model is applied to MEG data, leading to a novel approach to the investigation of functional connectivity.

In functional data analysis, data are commonly assumed to be smooth functions on a fixed interval of the real line. In this work, we introduce a comprehensive framework for the analysis of functional data, whose domain is a two-dimensional manifold and the domain itself is subject to variability from sample to sample. We formulate a statistical model for such data, here called functions on surfaces, which enables a joint representation of the geometric and functional aspects, and propose an associated estimation framework. We assess the validity of the framework by performing a simulation study and we finally apply it to the analysis of neuroimaging data of cortical thickness, acquired from the brains of different subjects, and thus lying on domains with different geometries. Supplementary materials for this article are available online.

Motivated by the analysis of high-dimensional neuroimaging signals located over the cortical surface, we introduce a novel Principal Component Analysis technique that can handle functional data located over a twodimensional manifold. For this purpose a regularization approach is adopted, introducing a smoothing penalty coherent with the geodesic distance over the manifold. The model introduced can be applied to any manifold topology, and can naturally handle missing data and functional samples evaluated in different grids of points. We approach the discretization task by means of finite element analysis, and propose an efficient iterative algorithm for its resolution. We compare the performances of the proposed algorithm with other approaches classically adopted in literature. We finally apply the proposed method to resting state functional magnetic resonance imaging data from the Human Connectome Project, where the method shows substantial differential variations between brain regions that were not apparent with other approaches.

BACKGROUND AND PURPOSE: Intracranial aneurysms (IAs), detected in 2%–5% of the population, represent a major health care issue because ruptured aneurysms with resultant hemorrhage are associated with severe morbidity or mortality. With the increasing role of artificial intelligence (AI) in diagnostic radiology, we assessed the accuracy of a commercial AI tool (Aidoc) to detect intracranial aneurysms on head or head/neck CTA relative to fellowship-trained neuroradiologists. MATERIALS AND METHODS: We retrospectively extracted CTA head or head/neck studies from University of Washington Medical Center’s clinical database between November 1, 2018 and November 2, 2021; these were analyzed with Aidoc for evaluation of aneurysm presence. Concordance or discrepancies between AI and the neuroradiology reports were further adjudicated by 3 neuroradiologists for consensus. IA features including size, morphology, and site of origin were extracted for each positive case. Correlation between AI and neuroradiologist performance was assessed, and a vascular neurosurgeon independently reviewed neuroradiology false-negatives to determine IA management based on image and patient-specific features. Comparative analyses were also performed per Aidoc’s intended use criteria, ie, “unruptured,” “saccular” IAs greater than 5 mm in size. RESULTS: A total of 2534 CTA scans were reviewed for IAs; 252 were positive with 315 IAs (1.25 aneurysms per positive CTA). AI achieved sensitivity, specificity, and accuracy of 70.5% (95% CI, 65.1%–75.5%), 98.6% (95% CI, 98.0%–99.0%), and 95.1% (95% CI, 94.3%–95.9%), respectively; while neuroradiologists’ performance were 94.0% (95% CI, 90.7%–96.3%), 98.3% (95% CI, 97.7%–98.8%), and 97.8% (95% CI, 97.1%–98.3%), respectively. In the cohort, 35 IAs were within the intended use criteria for Aidoc, and here AI achieved sensitivity, specificity, and accuracy of 85.7% (95% CI, 69.7%–95.2%), 98.6% (95% CI, 98.0%–99.0%), and 98.4% (95% CI, 97.8%–98.8%) while neuroradiologists achieved 97.1% (95% CI, 85.1%–99.9%), 98.3% (95% CI, 97.7%–98.8%), and 98.3% (95% CI, 97.7%–98.8%), respectively. CONCLUSIONS: Our multisite study showed that neuroradiologists performed better than AI for IA detection in terms of sensitivity and accuracy, while while achieving comparable specificity.

With the rapid progress of large language models (LLMs), advanced multimodal large language models (MLLMs) have demonstrated impressive zero-shot capabilities on vision-language tasks. In the biomedical domain, however, even state-of-the-art MLLMs struggle with basic Medical Decision Making (MDM) tasks. We investigate this limitation using two challenging datasets: (1) three-stage Alzheimer’s disease (AD) classification (normal, mild cognitive impairment, dementia), where category differences are visually subtle, and (2) MIMIC-CXR chest radiograph classification with 14 non-mutually exclusive conditions. Our empirical study shows that text-only reasoning consistently outperforms vision-only or vision-text settings, with multimodal inputs often performing worse than text alone. To mitigate this, we explore three strategies: (1) in-context learning with reason-annotated exemplars, (2) vision captioning followed by text-only inference, and (3) few-shot fine-tuning of the vision tower with classification supervision. These findings reveal that current MLLMs lack grounded visual understanding and point to promising directions for improving multimodal decision making in healthcare.

Background: Huntington’s disease is a progressive, autosomal dominant, neurodegenerative disease caused by a CAG repeat expansion in the HTT gene. Medium spiny neurons of the striatum are especially vulnerable to the disease, and atrophy of the caudate and putamen can be documented by neuroimaging years before the onset of symptoms. Objective: In this study, we aimed to characterize region-specific gray and white matter differences between Huntington’s disease patients and controls. Methods: We conducted a postmortem MRI study of the brains of 15 adults diagnosed with symptomatic Huntington’s disease and 26 control subjects, aiming to compare the differences in regional grey and white matter volumes between the two groups. Results: The study revealed decreased volumes in both grey and white matter in patients with Huntington’s disease, with the largest effect sizes observed in caudate and putamen. Notably, the atrophy predominantly affected the left hemisphere, particularly impacting grey and white matter regions adjacent to the pars opercularis, precentral, supramarginal, and pars orbitalis gyri, and the lateral orbitofrontal cortex. In the control group, asymmetry stems from larger left hemisphere regions compared to right, whereas an opposite pattern is observed in the Huntington’s disease group. Conclusions: These results suggest progressive, diffuse, and asymmetric grey and white matter atrophy occurs in Huntington’s disease. The reasons for this asymmetry remain unknown; however, our study provides a more detailed characterization of previously reported grey and white matter changes in Huntington’s disease, as observed through postmortem histopathological and MRI studies.

Purpose\enspace Embolic stroke of unidentified source (ESUS) represents 10–25% of all ischemic strokes. Our goal was to determine whether ESUS could be reclassified to cardioembolic (CE) or large-artery atherosclerosis (LAA) with machine learning (ML) using conventional clinical data. Methods\enspace We retrospectively collected conventional clinical features, including patient, imaging (MRI, CT/CTA), cardiac, and serum data from established cases of CE and LAA stroke, and factors with p\,< 0.2 in univariable analysis were used for creating a ML predictive tool. We then applied this tool to ESUS cases, with\,≥ 75% likelihood serving as the threshold for reclassification to CE or LAA. In patients with longitudinal data, we evaluated future cardiovascular events. Results\enspace 191 ischemic stroke patients (80 CE, 61 LAA, 50 ESUS) were included. Seven and 6 predictors positively associated with CE and LAA etiology, respectively. The c-statistic for discrimination between CE and LAA was 0.88. The strongest predictors for CE were left atrial volume index (OR\,= 2.17 per 1 SD increase) and BNP (OR\,= 1.83 per 1 SD increase), while the number of non-calcified stenoses\,≥ 30% upstream (OR\,= 0.34 per 1 SD increase) and not upstream (OR\,= 0.74 per 1 SD increase) from the infarct were for LAA. When applied to ESUS cases, the model reclassified 40% (20/50), with 11/50 reclassified to CE and 9/50 reclassified to LAA. In 21/50 ESUS with 30-day cardiac monitoring, 1/4 in CE and 3/16 equivocal reclassifications registered cardiac events, while 0/1 LAA reclassifications showed events. Conclusion\enspace ML tools built using standard ischemic stroke workup clinical biomarkers can potentially reclassify ESUS stroke patients into cardioembolic or atherosclerotic etiology categories.

\textexclamdown sec\textquestiondown\textexclamdown title\textquestiondown Introduction\textexclamdown/title\textquestiondown\textexclamdown p\textquestiondown The Human Connectome Project (HCP) has become a keystone dataset in human neuroscience, with a plethora of important applications in advancing brain imaging methods and an understanding of the human brain. We focused on tractometry of HCP diffusion-weighted MRI (dMRI) data.\textexclamdown/p\textquestiondown\textexclamdown/sec\textquestiondown\textexclamdown sec\textquestiondown\textexclamdown title\textquestiondown Methods\textexclamdown/title\textquestiondown\textexclamdown p\textquestiondown We used an open-source software library (pyAFQ; \textexclamdown ext-link ext-link-type="uri" xlink:href="https://yeatmanlab.github.io/pyAFQ" xmlns:xlink="http://www.w3.org/1999/xlink"\textquestiondown https://yeatmanlab.github.io/pyAFQ\textexclamdown/ext-link\textquestiondown ) to perform probabilistic tractography and delineate the major white matter pathways in the HCP subjects that have a complete dMRI acquisition (\textexclamdown italic\textquestiondown n\textexclamdown/italic\textquestiondown = 1,041). We used diffusion kurtosis imaging (DKI) to model white matter microstructure in each voxel of the white matter, and extracted tract profiles of DKI-derived tissue properties along the length of the tracts. We explored the empirical properties of the data: first, we assessed the heritability of DKI tissue properties using the known genetic linkage of the large number of twin pairs sampled in HCP. Second, we tested the ability of tractometry to serve as the basis for predictive models of individual characteristics (e.g., age, crystallized/fluid intelligence, reading ability, etc.), compared to local connectome features. To facilitate the exploration of the dataset we created a new web-based visualization tool and use this tool to visualize the data in the HCP tractometry dataset. Finally, we used the HCP dataset as a test-bed for a new technological innovation: the TRX file-format for representation of dMRI-based streamlines.\textexclamdown/p\textquestiondown\textexclamdown/sec\textquestiondown\textexclamdown sec\textquestiondown\textexclamdown title\textquestiondown Results\textexclamdown/title\textquestiondown\textexclamdown p\textquestiondown We released the processing outputs and tract profiles as a publicly available data resource through the AWS Open Data program’s Open Neurodata repository. We found heritability as high as 0.9 for DKI-based metrics in some brain pathways. We also found that tractometry extracts as much useful information about individual differences as the local connectome method. We released a new web-based visualization tool for tractometry—“Tractoscope” (\textexclamdown ext-link ext-link-type="uri" xlink:href="https://nrdg.github.io/tractoscope" xmlns:xlink="http://www.w3.org/1999/xlink"\textquestiondown https://nrdg.github.io/tractoscope\textexclamdown/ext-link\textquestiondown ). We found that the TRX files require considerably less disk space-a crucial attribute for large datasets like HCP. In addition, TRX incorporates a specification for grouping streamlines, further simplifying tractometry analysis.\textexclamdown/p\textquestiondown\textexclamdown/sec\textquestiondown