Uncertainty-Aware Foundation Models for Clinical Data

Forum topic · 小凯 · 2026-04-07

Summary

This paper proposes an uncertainty-aware foundation model framework for clinical data. Instead of representing each patient as a single point embedding, the model represents patients as distributions over plausible latent states, capturing what is invariantly inferable while explicitly encoding epistemic uncertainty. The framework uses set-valued representations with consistency enforced across partial views of the same patient, integrated with multimodal encoders and scalable self-supervised objectives combining reconstruction, contrastive alignment, and distributional regularization. Experiments across diverse clinical tasks show improvements in predictive performance, robustness under missing data, and uncertainty calibration compared to strong baselines. The authors argue that modeling unobserved information—rather than only observed data—serves as a critical inductive bias for healthcare foundation models, addressing the incompleteness inherent in sparse, irregular, modality-dependent clinical measurements.

Overview

Field: Machine Learning
Authors: Qian Zhou, Yuanyun Zhang, Shi Li

Abstract

Healthcare foundation models have largely followed paradigms from natural language processing and computer vision, emphasizing large scale pretraining and deterministic representations over heterogeneous clinical data. However, clinical observations are inherently incomplete, reflecting sparse, irregular, and modality dependent measurements of an underlying physiologic state. In this work, we propose a framework for uncertainty aware foundation modeling that represents each patient not as a point embedding, but as a distribution over plausible latent states. By learning set valued representations and enforcing consistency across partial views of the same patient, the model captures what is invariantly inferable while explicitly encoding epistemic uncertainty. We integrate this formulation with multimodal encoders and scalable self supervised objectives, combining reconstruction, contrastive alignment, and distributional regularization. Across diverse clinical tasks, our approach improves predictive performance, robustness under missing data, and uncertainty calibration relative to strong baselines. These results suggest that modeling what is not observed rather than only what is constitutes a critical inductive bias for healthcare foundation models.

Key Ideas

Distributional patient representation: Each patient is modeled as a distribution over plausible latent states rather than a deterministic point embedding.
Set-valued representations with partial-view consistency: Consistency across partial views of the same patient captures invariantly inferable information and encodes epistemic uncertainty explicitly.
Multimodal, self-supervised training: Combines reconstruction, contrastive alignment, and distributional regularization objectives at scale.

Results

Improved predictive performance across diverse clinical tasks.
Greater robustness under missing data.
Better uncertainty calibration relative to strong baselines.