Learning Dynamic Representations and Policies from Multimodal Clinical Time-Series with Informative Missingness

2026-04-22

AI Summary

This study addresses the challenge of informative missingness in clinical multimodal time-series data, where conventional methods often overlook the latent patient states encoded in missingness patterns. The authors propose a novel framework that explicitly models informative missingness by integrating structured measurements, clinical notes, and their observation patterns into a unified multimodal encoder. Leveraging Bayesian filtering, the model dynamically updates the patient’s latent state over time, treating the observation process itself as a meaningful signal. This approach enables improved treatment policy evaluation and outcome prediction within an offline reinforcement learning setting. Evaluated on the MIMIC-III sepsis cohort, the method achieves state-of-the-art performance, yielding a Fitted Q-Evaluation (FQE) score of 0.679 for treatment policy assessment and an AUROC of 0.886 for 72-hour mortality prediction.

Technology Category

• Machine Learning: Multimodal Learning
• Reasoning under Uncertainty: Relational Probabilistic Models
• Computer Vision: Multi-modal Vision

Application Category

• User Modeling, Personalization and Recommendation: Psychology-informed user models and recommender systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs

Abstract

Multimodal clinical records contain structured measurements and clinical notes recorded over time, offering rich temporal information about the evolution of patient health. Yet these observations are sparse, and whether they are recorded depends on the patient's latent condition. Observation patterns also differ across modalities, as structured measurements and clinical notes arise under distinct recording processes. While prior work has developed methods that accommodate missingness in clinical time series, how to extract and use the information carried by the observation process itself remains underexplored. We therefore propose a patient representation learning framework for multimodal clinical time series that explicitly leverages informative missingness. The framework combines (1) a multimodal encoder that captures signals from structured and textual data together with their observation patterns, (2) a Bayesian filtering module that updates a latent patient state over time from observed multimodal signals, and (3) downstream modules for offline treatment policy learning and patient outcome prediction based on the learned patient state. We evaluate the framework on ICU sepsis cohorts from MIMIC-III, MIMIC-IV, and eICU. It improves both offline treatment policy learning and adverse outcome prediction, achieving FQE 0.679 versus 0.528 for clinician behavior and AUROC 0.886 for post-72-hour mortality prediction on MIMIC-III.

Problem

• multimodal clinical time-series
• informative missingness
• patient representation learning
• observation process
• dynamic representations

Innovation

• informative missingness
• multimodal clinical time series
• Bayesian filtering
• patient representation learning
• offline policy learning