Embracing Aleatoric Uncertainty in Medical Multimodal Learning with Missing Modalities
Linxiao Gong, Yang Liu, Lianlong Sun, Yulai Bi, Jing Liu, Xiaoguang Zhu
TL;DR
The paper tackles missing modalities in medical multimodal learning by explicitly modeling aleatoric uncertainty. It introduces Aleatoric Uncertainty Modeling (AUM), representing each unimodal input as a Gaussian with learned mean $\mu_i^m$ and variance $(\sigma_i^m)^2$, and fuses modalities through an uncertainty-aware dynamic bipartite graph. AUM also incorporates a variational information bottleneck loss and uncertainty-guided attention, enabling natural handling of missing data without reconstruction and achieving mortality-prediction gains of 2.26% (MIMIC-IV) and 2.17% (eICU) in AUC-ROC. This approach advances robust clinical decision support by providing explicit uncertainty quantification and a scalable fusion mechanism for incomplete multimodal data.
Abstract
Medical multimodal learning faces significant challenges with missing modalities prevalent in clinical practice. Existing approaches assume equal contribution of modality and random missing patterns, neglecting inherent uncertainty in medical data acquisition. In this regard, we propose the Aleatoric Uncertainty Modeling (AUM) that explicitly quantifies unimodal aleatoric uncertainty to address missing modalities. Specifically, AUM models each unimodal representation as a multivariate Gaussian distribution to capture aleatoric uncertainty and enable principled modality reliability quantification. To adaptively aggregate captured information, we develop a dynamic message-passing mechanism within a bipartite patient-modality graph using uncertainty-aware aggregation mechanism. Through this process, missing modalities are naturally accommodated, while more reliable information from available modalities is dynamically emphasized to guide representation generation. Our AUM framework achieves an improvement of 2.26% AUC-ROC on MIMIC-IV mortality prediction and 2.17% gain on eICU, outperforming existing state-of-the-art approaches.