Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bayesian Integration of Nonlinear Incomplete Clinical Data

Lucía González-Zamorano, Nuria Balbás-Esteban, Vanessa Gómez-Verdejo, Albert Belenguer-Llorens, Carlos Sevilla-Salcedo

TL;DR

BIONIC tackles the challenge of integrating heterogeneous, high-dimensional clinical data with view-level missingness by learning a Bayesian dual-latent model that combines a generative space for missing data with a discriminative space for prediction. It uses pretrained, fixed embeddings as multimodal views and imposes sparsity through ARD priors to automatically prune irrelevant factors, while a joint inference scheme handles missing views and labels in a semi-supervised setting. The work contributes a principled treatment of missingness across variables, views, and labels, an intrinsic interpretability pipeline linking latent factors to inputs, and empirical validation across three diverse clinical datasets showing robust discriminative performance and improved calibration under incomplete data. The approach offers practical advantages for clinical decision support by enabling principled imputation, uncertainty quantification, and population-level insights into modality relevance, with potential extensions to temporal data and patient-specific explanations.

Abstract

Multimodal clinical data are characterized by high dimensionality, heterogeneous representations, and structured missingness, posing significant challenges for predictive modeling, data integration, and interpretability. We propose BIONIC (Bayesian Integration of Nonlinear Incomplete Clinical data), a unified probabilistic framework that integrates heterogeneous multimodal data under missingness through a joint generative-discriminative latent architecture. BIONIC uses pretrained embeddings for complex modalities such as medical images and clinical text, while incorporating structured clinical variables directly within a Bayesian multimodal formulation. The proposed framework enables robust learning in partially observed and semi-supervised settings by explicitly modeling modality-level and variable-level missingness, as well as missing labels. We evaluate BIONIC on three multimodal clinical and biomedical datasets, demonstrating strong and consistent discriminative performance compared to representative multimodal baselines, particularly under incomplete data scenarios. Beyond predictive accuracy, BIONIC provides intrinsic interpretability through its latent structure, enabling population-level analysis of modality relevance and supporting clinically meaningful insight.

Bayesian Integration of Nonlinear Incomplete Clinical Data

TL;DR

BIONIC tackles the challenge of integrating heterogeneous, high-dimensional clinical data with view-level missingness by learning a Bayesian dual-latent model that combines a generative space for missing data with a discriminative space for prediction. It uses pretrained, fixed embeddings as multimodal views and imposes sparsity through ARD priors to automatically prune irrelevant factors, while a joint inference scheme handles missing views and labels in a semi-supervised setting. The work contributes a principled treatment of missingness across variables, views, and labels, an intrinsic interpretability pipeline linking latent factors to inputs, and empirical validation across three diverse clinical datasets showing robust discriminative performance and improved calibration under incomplete data. The approach offers practical advantages for clinical decision support by enabling principled imputation, uncertainty quantification, and population-level insights into modality relevance, with potential extensions to temporal data and patient-specific explanations.

Abstract

Multimodal clinical data are characterized by high dimensionality, heterogeneous representations, and structured missingness, posing significant challenges for predictive modeling, data integration, and interpretability. We propose BIONIC (Bayesian Integration of Nonlinear Incomplete Clinical data), a unified probabilistic framework that integrates heterogeneous multimodal data under missingness through a joint generative-discriminative latent architecture. BIONIC uses pretrained embeddings for complex modalities such as medical images and clinical text, while incorporating structured clinical variables directly within a Bayesian multimodal formulation. The proposed framework enables robust learning in partially observed and semi-supervised settings by explicitly modeling modality-level and variable-level missingness, as well as missing labels. We evaluate BIONIC on three multimodal clinical and biomedical datasets, demonstrating strong and consistent discriminative performance compared to representative multimodal baselines, particularly under incomplete data scenarios. Beyond predictive accuracy, BIONIC provides intrinsic interpretability through its latent structure, enabling population-level analysis of modality relevance and supporting clinically meaningful insight.
Paper Structure (14 sections, 5 equations, 3 figures, 4 tables)

This paper contains 14 sections, 5 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Probabilistic modeling and inference
  3. Pretrained embeddings as multimodal views
  4. Interpretability pipeline
  5. Materials
  6. MMIST
  7. MOTUM
  8. TCGA
  9. Results and Discussion
  10. Experimental Setup
  11. Discriminative Performance
  12. Generative Performance
  13. Interpretability Analysis
  14. Conclusions

Figures (3)

  • Figure 1: Overview of the proposed BIONIC framework. Structured clinical variables and pretrained embeddings are mapped through view-specific projections into a dual latent space, where a generative component models multimodal structure and missing data, and a discriminative component supports the downstream prediction task.
  • Figure 2: Voxel-wise relevance map obtained via sensitivity analysis, overlaid onto a representative reconstructed patient scan (axial view). Positive intensity changes (red) are associated with primary tumor and negative values (blue) are associated with metastasis.
  • Figure 3: Token-level discriminative relevance for representative non-basal (top) and basal (bottom) patients. Red denotes positive relevance and blue negative relevance, with color intensity proportional to relevance magnitude.