Brain pathway anchored multimodal generative representations for patient-specific predictions of Parkinson’s disease

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: long paper (4–8 pages excluding references)

Keywords: Parkinson’s disease, Multimodal neuroimaging, Variational co-clustering, Interpretable machine learning

TL;DR: A pathway-anchored stratification strategy by summarizing multimodal features within each brain circuit using a Multimodal Pathway Integrity Scores (MPIS).

Abstract: Parkinson's disease is increasingly understood as a disorder of distributed brain circuits, yet most imaging analyses do not explicitly respect pathway structure. We introduce a pathway-anchored, multimodal clustering framework based on Scalable Robust Variational Compositional Co-clustering (SRVCC) that integrates structural MRI, diffusion MRI, and DAT-SPECT in anatomically defined circuits. For each pathway, we derive a simple Multimodal Pathway Integrity Score (MPIS) that aggregates $z$-normalised volume, microstructural, and dopaminergic measures into an interpretable summary of imaging integrity. Motivated by the need for patient sub-subtyping and improved diagnostic specificity, we develop these new generative, pathway-aware representations to capture circuit-level heterogeneity that may be obscured by region- or modality-centric analyses. In the PPMI cohort, SRVCC identifies stable imaging-derived patient clusters and feature modules under explicit model selection and bootstrap/stability checks, with covariate adjusted analyzes controlling for age, sex, education, and medication. MPIS shows coherent structure function associations. Feature-level reports highlight dominant region by modality contributors, providing a transparent bridge from multimodal data to circuit-level signatures. This pathway-aware representation offers a principled, reproducible way to summarize multimodal imaging in PD and may support future work on circuit-informed stratification, prognosis, and targeted outcome measures, helping clinicians deliver more specific diagnoses and better-tailored interventions toward precision healthcare.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Chandrajit_L._Bajaj1

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 66