AACE: Annotation-Assisted Learning of Treatment Policies from Multimodal EHRs
AACE (Annotation-Assisted Causal policy lEarning) has been introduced by researchers as a technique for deriving treatment policies from multimodal electronic health records (EHRs), which integrate both tabular data and clinical narratives. Current causal policy learning techniques primarily focus on tabular variables and often struggle with multimodal data. Although predictive models assessing baseline risk are prevalent, they are not tailored to pinpoint which patients gain the most from treatment. AACE enhances representation learning through annotations, aiming to minimize bias in treatment effect evaluations. This innovative method aspires to assist healthcare providers in making informed treatment choices and optimizing the allocation of healthcare resources.
Key facts
- AACE stands for Annotation-Assisted Causal policy lEarning.
- The method learns treatment policies from multimodal EHRs (tabular data + clinical text).
- Existing causal estimators are designed for tabular covariates and may not preserve confounding information in multimodal settings.
- Predictive models of baseline risk are commonly used but not designed to identify patients with largest expected treatment benefit.
- AACE uses annotations to improve representation learning and reduce bias.
- The goal is to help physicians make better treatment decisions and allocate healthcare resources efficiently.
- The paper is available on arXiv with ID 2507.20993.
- The announcement type is replace-cross.
Entities
Institutions
- arXiv