Causal Debiasing Medical Multimodal Representation Learning with Missing Modalities

2025-09-06

AI Summary

Real-world medical multimodal data often exhibit non-random modality missingness due to cost or clinical constraints, inducing two types of bias: *missingness bias*—arising from non-ignorable missing mechanisms—and *distributional bias*—driven by confounding factors. To address these, we propose the first structural causal model (SCM)-based dual debiasing framework. It employs backdoor adjustment to rigorously identify and eliminate both biases. Our method introduces a missingness deconfounding module and a dual-branch network that disentangles causal features from spurious correlations, enabling end-to-end training. Extensive experiments on multiple real-world public and in-hospital multimodal datasets demonstrate substantial improvements in predictive performance under incomplete modalities. Moreover, the learned representations are inherently interpretable and causally grounded.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multimodal Learning
• Reasoning under Uncertainty: Causality

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Medical multimodal representation learning aims to integrate heterogeneous clinical data into unified patient representations to support predictive modeling, which remains an essential yet challenging task in the medical data mining community. However, real-world medical datasets often suffer from missing modalities due to cost, protocol, or patient-specific constraints. Existing methods primarily address this issue by learning from the available observations in either the raw data space or feature space, but typically neglect the underlying bias introduced by the data acquisition process itself. In this work, we identify two types of biases that hinder model generalization: missingness bias, which results from non-random patterns in modality availability, and distribution bias, which arises from latent confounders that influence both observed features and outcomes. To address these challenges, we perform a structural causal analysis of the data-generating process and propose a unified framework that is compatible with existing direct prediction-based multimodal learning methods. Our method consists of two key components: (1) a missingness deconfounding module that approximates causal intervention based on backdoor adjustment and (2) a dual-branch neural network that explicitly disentangles causal features from spurious correlations. We evaluated our method in real-world public and in-hospital datasets, demonstrating its effectiveness and causal insights.

Problem

• Addressing missingness bias in medical multimodal data
• Mitigating distribution bias from latent confounders
• Developing causal debiasing framework for incomplete modalities

Innovation

• Causal intervention via backdoor adjustment
• Dual-branch network disentangles causal features
• Structural analysis addresses missingness and distribution biases