A Representation Learning Framework for Clinical Trajectories from Multimodal Longitudinal EHRs

ACL ARR 2026 January Submission9655 Authors

Keywords: Clinical Trajectory, Clinical Notes, Contrastive Question Answering, Temporal Modeling

Abstract: Understanding heterogeneity in disease progression from longitudinal electronic health records (EHRs) remains challenging due to irregular temporal sampling, missing data, and unstructured clinical narratives. We present an empirical framework that integrates QA-based abstraction of clinical notes with multimodal temporal representation learning to study progression-aware patient representations. By aligning structured physiological measurements and interpretable semantic signals derived from clinical narratives, the framework induces admission-level trajectory representations under temporal irregularity. Using a sepsis cohort from MIMIC-III, we analyze latent trajectory groupings in the learned representation space and characterize their progression patterns, outcomes, and early predictability. Our findings suggest that QA-based abstraction of clinical text, when combined with temporal multimodal modeling, yields more coherent and clinically aligned trajectory representations than existing baselines. The code and data are available at \url{https://github.com/anonymous-2344/clinical_trajectory_phenotype}.

Paper Type: Long

Research Area: Clinical and Biomedical Applications

Research Area Keywords: Clinical and Biomedical Applications, Information Extraction, Information Retrieval and Text Mining, Language Modeling, NLP Applications, Question Answering

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 9655