Learning from Multimodal Informative Missingness
Background: In multimodal clinical time-series, the observation processes for structured measurements and textual notes are often informative, reflecting underlying patient severity and clinician behavior.
Question / Future Work: It remains open how to effectively leverage informative missingness over time across multiple modalities to learn dynamic patient representations suitable for sequential clinical decision-making. Future research is needed to determine the optimal integration of evolving observation patterns, such as documentation burstiness and measurement gaps, to improve both the accuracy of health state representation and the robustness of learned treatment policies.
Why It Matters: This is a core challenge in clinical machine learning, as most existing models treat missing data merely as a technical problem to be imputed, rather than as a source of clinical signal.
Evidence: However, it remains open how to leverage informative missingness over time across modalities to learn patient representations.
Metadata & Links
- created_at
- 2026-04-26T05:52:06Z