Multi-View Variational Autoencoders Allow for Interpretability Leveraging Digital Avatars: Application to the HBN Cohort

Corentin Ambroise, Antoine Grigis, Édouard Duchesnay, Vincent Frouin2023 · DOI

Abstract

While neural network-based methods are praised for their prediction performance, they are often criticized for their lack of interpretability. When dealing with multi-modal data, models must learn the independent and joint effects of heterogeneous views while yielding interpretable results. In this work, we demonstrate how multi-view variational autoencoders (VAEs) provide a convenient statistical framework to learn the input data joint distribution. We design a method that discovers relationships between views by generating "digital avatars"—synthetic brain representations corresponding to specific clinical statuses. By computing associations between cortical regions and behavioral scores in the Healthy Brain Network (HBN) cohort, we showcase that such digital avatars convey interpretable information for children experiencing mental health issues.

Introduction

In the study of complex conditions like Autism Spectrum Disorder (ASD), researchers manage multi-modal data from diverse sources, including MRI scans and clinical questionnaires. While deep learning models excel at pattern recognition, they are notoriously difficult to interpret. Traditional ASD research often attempts to force patients into rigid subgroups that fail to replicate across datasets. This paper argues for a "dimensional approach" that examines the entire spectrum of variability. The challenge lies in building a model that integrates diverse data types while remaining transparent enough for clinicians to understand underlying biological relationships.

Methodology: The Digital Avatar Framework

The proposed framework utilizes the MoPoe-VAE (Mixture-of-Product-of-Experts Variational Autoencoder) architecture. Unlike standard VAEs, this model handles multiple "views" (e.g., imaging and clinical scores) and remains robust even with missing data. It learns a joint latent space that captures the essence of both brain structure and behavior.

The Digital Avatar framework: a three-step pipeline involving inference, simulation via clinical score perturbation, and statistical association.

The core innovation is a three-step interpretability pipeline:

  1. Inference: The model maps a subject into the joint latent space.
  2. Simulation: Researchers "perturb" a specific clinical score (e.g., increasing an anxiety score) and use the model to generate what that child's brain would look like at that new status. This synthetic reconstruction is the "digital avatar."
  3. Association: By repeating this for many subjects and using hierarchical linear regressions, the authors statistically map which brain regions are most sensitive to changes in specific behaviors.

Experimental Setup and Datasets

The study utilizes the Healthy Brain Network (HBN) cohort, focusing on psychiatric disorders in children and adolescents. The dataset includes 2,991 samples, with 1,505 subjects providing both imaging and clinical data.

  • Clinical View: Seven scores assessing social responsiveness (SRS), anxiety (SCARED), depression (ARI), and behavioral disorders (CBCL).
  • Image View: Structural MRI data processed to extract cortical thickness, curvature, and area across the Destrieux parcellation.

Key Results and Discussion

The model successfully isolated biological signals from technical noise. Representational Similarity Analysis (RSA) revealed that while the "Image" latent space encoded clinical site effects, the "Joint" latent space was significantly less impacted.

Radar plots and brain maps showing the associations between clinical scores (SRS, SCARED, etc.) and cortical thickness, curvature, and area.

The framework identified several significant associations aligning with ASD literature:

  • Social Responsiveness (SRS): Higher challenges were associated with increased cortical thickness in the right insula and left temporal superior regions, but decreased thickness in the central sulcus.
  • Anxiety and Seclusion: Distinct patterns were found; for instance, left and right temporal areas were linked to anxiety and seclusion scores through curvature and area measurements.
  • Statistical Robustness: Results were validated using bootstrapping (reproducible in >70% of experiments) and Bonferroni correction.

Conclusion

By moving beyond rigid subtyping, this framework treats ASD as a continuous spectrum. While currently limited by one-at-a-time feature perturbation, future work will explore multi-feature interactions and longitudinal data to better reflect biological reality and track symptom evolution over time.

Summary updated

Abstract

If neural network-based methods are praised for their prediction performance, they are often criticized for their lack of interpretability. When dealing with multi-omics or multi-modal data, neural network methods must be able learn the independent and joint effect of heterogeneous views while yielding interpretable results intra- and interviews. In the literature, multi-view generative models exist to learn joint information in a reduced-size latent space. Among these models, multi-view variational autoencoders are very promising. In this work, we demonstrate how they provide a convenient statistical framework to learn the input data joint distribution and offer opportunities for the results interpretation. We design a method that discovers the relationships between one view and others. The generative capabilities of the model enable the exploration of a whole disorder spectrum through the generation of realistic values. While modifying a subject’s clinical score, the model retrieves a representation of the subject’s brain at this clinical status, so-called digital avatar. By computing associations between cortical regions measures and behavioral scores, we showcase that such digital avatars convey interpretable information in a multi-modal cohort with children experiencing mental health issues.
BibTeX
@misc{ambroise2023multi,
  author = {Corentin Ambroise and Antoine Grigis and Édouard Duchesnay and Vincent Frouin},
  title = {{Multi-View Variational Autoencoders Allow for Interpretability Leveraging Digital Avatars: Application to the HBN Cohort}},
  year = {2023},
  doi = {10.1109/isbi53787.2023.10230552},
}
Figure