arXiv 2509.056152025-09-06cs.LGcs.AI

缺失模态下的医学多模态表示学习因果去偏

Causal Debiasing Medical Multimodal Representation Learning with Missing Modalities

DataLab: Data Science and Informatics, University of California, Davis（加州大学戴维斯分校数据实验室）；Department of Electrical and Computer Engineering, University of Rochester（罗切斯特大学电气与计算机工程系）；Academy for Engineering & Technology, Fudan University（复旦大学工程与技术学院）；Department of Computer Science, University of Toronto（多伦多大学计算机科学系）；Department of Electrical and Computer Engineering, New York University（纽约大学电气与计算机工程系）；UC Davis Health（加州大学戴维斯分校医疗中心）；Department of Basic Medical Sciences, University of Arizona（亚利桑那大学基础医学系）；Department of Computer Science, University of California, Davis（加州大学戴维斯分校计算机科学系）

Xiaoguang Zhu, Lianlong Sun, Yang Liu, Pengyi Jiang, Uma Srivatsa, Nipavan Chiamvimonvat, Vladimir Filkov

备注：: Submitted to IEEE TKDE

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

针对医学多模态学习中模态缺失引发的偏差，提出基于因果分析和后门调整的去偏框架，有效提升模型泛化能力。

摘要：

医学多模态表示学习旨在将异质性临床数据整合为统一的患者表示以支持预测建模，这在医学数据挖掘社区中仍是一项重要且具有挑战性的任务。然而，由于成本、协议或患者特定限制，现实世界的医学数据集常常遭受模态缺失问题。现有方法主要通过从原始数据空间或特征空间中的可用观测中学习来解决此问题，但通常忽略了数据采集过程本身引入的潜在偏差。在本工作中，我们识别出阻碍模型泛化的两类偏差：缺失性偏差（由模态可用性的非随机模式导致）和分布性偏差（由影响观测特征和结果的潜在混杂因素引起）。为应对这些挑战，我们对数据生成过程进行了结构因果分析，并提出一个与现有基于直接预测的多模态学习方法兼容的统一框架。我们的方法包含两个关键组件：（1）一个基于后门调整近似因果干预的缺失性去混杂模块，以及（2）一个显式将因果特征与虚假相关性分离的双分支神经网络。我们在真实世界的公开数据集和院内数据集上评估了我们的方法，证明了其有效性和因果洞察力。

英文摘要：

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.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。