Unveiling Patient Group Dynamics through Multimodal Manifold Trajectories
Poster No:
1615
Submission Type:
Abstract Submission
Authors:
Hooman Rokham1, Haleh Falakshahi1, Vince Calhoun2
Institutions:
1Georgia State University, Atlanta, GA, 2GSU/GATech/Emory, Atlanta, GA
First Author:
Co-Author(s):
Introduction:
Understanding the complex interplay of mental health disorders requires innovative approaches that bridge the heterogeneity of patient data. Disorders like schizophrenia, bipolar disorder, and schizoaffective disorder exhibit overlapping yet distinct patterns of brain dysfunction. This study introduces a novel multimodal framework that fuses structural MRI (sMRI) and functional MRI (fMRI)-derived functional network connectivity (FNC) data to create a smooth, continuous 2D manifold. This manifold provides an interpretable visualization of patient group dynamics, offering insights into subtle transitions and unique trajectories among four groups: schizophrenia, bipolar disorder, schizoaffective disorder, and healthy controls.
Methods:
The methodology involves several steps: preprocessing, dimensionality reduction, multimodal feature fusion, and trajectory smoothing using t-distributed stochastic neighbor embedding (t-SNE). After preprocessing steps, Intrinsic connectivity networks (ICNs) were extracted using the fully automated NeuroMark independent component analysis (ICA) pipeline [1]. Next, high-dimensional neuroimaging data is vectorized and reduced to key features, retaining clinically relevant information. These features are then concatenated across modalities to construct a unified representation of each patient. The resulting dataset is mapped onto a 2D manifold using uniform manifold approximation and projection (UMAP), producing trajectories that capture local and global relationships across patient groups. Smooth trajectories are extracted by applying density estimation and Gaussian smoothing, enabling the identification of unique and overlapping pathways within the manifold. Figure 1 illustrates the workflow of the proposed method, including preprocessing, feature selection, multimodal fusion, UMAP application, and trajectory smoothing.
·Methodology Diagram illustrating preprocessing, feature selection, multimodal fusion, UMAP embedding, and trajectory smoothing.
Results:
Our results reveal distinct patterns within the 2D manifold, with schizophrenia and healthy controls occupying endpoints of the spectrum. Figure 2 visualizes the estimated 4-class trajectory, NC-SAD-BP-SZ (control, schizoaffective , bipolar, schizophrenia), demonstrating the continuum-like nature of these disorders. Additionally, it highlights the changes in FNC features derived from fMRI data and gray matter alterations observed in sMRI data along the path. For example, the control group shows higher subcortical and cerebellar connectivity compared to schizophrenia, while intermediate groups exhibit a mixed connectivity profile. Gray matter volume progressively changes along the trajectory, underscoring neuroanatomical differences across the spectrum.
These findings underscore the potential of trajectory-based modeling for advancing mental health research. By capturing the spectrum-like transitions between disorders, this framework moves beyond rigid classification systems and offers a biologically plausible representation of psychiatric conditions. Such insights can inform more personalized approaches to diagnosis and treatment, particularly in distinguishing between overlapping symptom profiles.
These findings underscore the potential of trajectory-based modeling for advancing mental health research. By capturing the spectrum-like transitions between disorders, this framework moves beyond rigid classification systems and offers a biologically plausible representation of psychiatric conditions. Such insights can inform more personalized approaches to diagnosis and treatment, particularly in distinguishing between overlapping symptom profiles.
·Visualization of the 4-class trajectory (NC-SAD-BP-SZ) on the 2D manifold. Displays FNC feature changes from fMRI and gray matter alterations from sMRI along the trajectory, highlighting transitions a
Conclusions:
In summary, the proposed multimodal framework integrates diverse neuroimaging data into a smooth, interpretable 2D manifold. This approach not only highlights the shared and distinct pathways among psychiatric disorders but also provides a foundation for understanding the continuum of mental health conditions. Future directions include extending this framework to longitudinal datasets and integrating additional modalities to enhance its clinical applicability.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 2
Multivariate Approaches 1
Keywords:
FUNCTIONAL MRI
Modeling
Schizophrenia
STRUCTURAL MRI
Other - multimodal fusion, trajectory mapping
1|2Indicates the priority used for review