Annotation-Assisted Learning of Treatment Policies From Multimodal Electronic Health Records

Henri Arno, Thomas DemeesterarXiv 2025 · arXiv

TL;DR

The paper introduces AACE (Annotation-Assisted Coarsened Effects), a novel framework for learning optimal treatment policies from multimodal electronic health records (EHRs) containing both tabular data and clinical text. By leveraging expert annotations only during training to adjust for hidden confounding factors in text, AACE accurately estimates treatment benefits using standard multimodal representations at inference time. It consistently outperforms traditional risk-based and representation-based causal baselines across synthetic, semi-synthetic, and real-world clinical datasets.

What Problem Does This Paper Address?

When medical resources are limited, doctors must prioritize treatments for patients who will benefit the most. Machine learning can help by learning "treatment policies" from observational Electronic Health Records (EHRs).

However, modern EHRs are multimodal, combining structured tabular data (like age or lab results) with unstructured clinical text (like physician notes). This creates a major hurdle for causal inference: unconfoundedness. Standard causal models assume that all variables influencing both the treatment decision and the patient outcome are perfectly measured and observed.

Overview of the problem setting, where key confounders may be trapped inside unstructured clinical text.

If a critical symptom (e.g., a severe cough) is only mentioned in the text, standard text embeddings might not fully preserve this confounding information. Directly applying causal estimators to these imperfect representations leads to biased treatment effect estimates. To bypass this, clinicians often fall back on predicting baseline risk (who is sickest), but being the sickest doesn't automatically mean a patient will benefit the most from a specific treatment.

Methods

The authors propose AACE (Annotation-Assisted Coarsened Effects), which treats expert annotations as privileged information—available during training to eliminate confounding bias, but completely unnecessary during deployment.

AACE operates in three distinct stages:

  1. Pseudo-Outcome Construction (Training): Using the full set of training covariates $X$—which includes both structured tabular data and explicit expert annotations of text-based confounders—the model estimates the propensity scores and response surfaces. It uses these to construct doubly robust (DR) pseudo-outcomes, effectively de-biasing the causal signal.
  2. Effect Estimation (Training): The model regresses these de-biased pseudo-outcomes against the standard multimodal representations $\phi$ (the text embeddings concatenated with tabular data). This teaches the model to predict the coarsened treatment effect based only on information available at inference time.
  3. Inference: At deployment, the expert annotator is removed from the loop. The model takes standard multimodal EHR data, encodes it, and directly predicts the expected treatment benefit to prioritize the top-$k$ patients.
The three-stage workflow of the AACE framework.

Datasets

The framework was evaluated across three distinct types of EHR datasets:

  • SynSum: A fully synthetic dataset of 10,000 records where clinical notes were generated using GPT-4. It features explicit text-based symptom confounders (like fever and dyspnea) with known ground-truth treatment effects.
  • MIMIC-Syn: A semi-synthetic benchmark built on real ICU notes and demographics from the MIMIC-III database. Treatments and outcomes were simulated using latent cardiovascular diagnoses that are implicitly expressed in the text.
  • MIMIC-Real: A fully real-world observational dataset from MIMIC-III tracking antibiotic administration and in-hospital mortality. Pre-treatment vital signs were utilized as the training-time covariates to adjust for confounding.

Key Insights and Ideas

  • The "Coarsening" Justification: The authors prove mathematically (Proposition 1) that even if text embeddings lose some confounding details (coarsening the effect), the resulting treatment policy can still be 100% optimal as long as the coarsening bias doesn't swap patients across the final treatment decision boundary.
  • Privileged Supervision: Instead of hoping that deep text encoders perfectly preserve causal structures, AACE uses targeted human expertise during the training phase to anchor the causal adjustments, making deep learning models safer for clinical policy design.

Key Results

AACE was evaluated against risk-based models and representation-based causal models (like TARNet and DragonNet) using metrics such as AUTOC (Area Under the Treatment-Outcome Curve) and PEHE (Precision in Estimating Heterogeneous Effects).

  • Superior Policy Value: On the SynSum dataset (with 4,000 training samples), AACE achieved an AUTOC of -0.56, significantly outperforming the risk baseline (-0.22) and representation-based causal baselines like DragonNet (-0.40).
  • Graceful Degradation: As shown in the paper's ablation studies, AACE's treatment effect estimation error (PEHE) steadily decreases as training sizes scale, outperforming tabular-only models by successfully extracting latent signals from text representations.
Ablation study on SynSum showing PEHE error decreasing as training size scales.
  • Real-World Divergence: On the real-world MIMIC-Real dataset, AACE and risk-based models showed incredibly low Jaccard overlap (e.g., only 3.7% agreement on the top 5% highest-priority patients). This highlights that the choice of modeling strategy drastically alters who actually receives care in practice.

Limitations and Future Work

While powerful, AACE does not prescribe a universal annotation protocol; the choice of what to annotate remains highly application-dependent and requires upfront domain expertise. Additionally, the quantitative evaluations relied on simulated or proxy annotations rather than real-world human-in-the-loop clinical labeling.

The authors suggest that future work should explore scaling up AACE using LLM-assisted labeling with expert verification, as well as validating the framework directly on Randomized Controlled Trial (RCT) data.

Summary updated

Abstract

We study how to learn treatment policies from multimodal electronic health records (EHRs) that consist of tabular data and clinical text. These policies can help physicians make better treatment decisions and allocate healthcare resources more efficiently. Causal policy learning methods prioritize patients with the largest expected treatment benefit. Yet, existing estimators are designed for tabular covariates under causal assumptions that may be hard to justify in the multimodal setting. A pragmatic alternative is to apply causal estimators directly to multimodal representations, but this can produce biased treatment effect estimates when the representations do not preserve the relevant confounding information. As a result, predictive models of baseline risk are commonly used in practice to guide treatment decisions, although they are not designed to identify which patients benefit most from treatment. We propose AACE (Annotation-Assisted Coarsened Effects), an annotation-assisted approach to causal policy learning for multimodal EHRs. The method uses expert-provided annotations during training to support confounding adjustment, and then predicts treatment benefit from only multimodal representations at inference. We show that the proposed method achieves strong empirical performance across synthetic, semi-synthetic, and real-world EHR datasets, outperforming risk-based and representation-based causal baselines, and offering practical insights for applying causal machine learning in clinical practice.
BibTeX
@misc{arno2025annotation,
  author = {Henri Arno and Thomas Demeester},
  title = {{Annotation-Assisted Learning of Treatment Policies From Multimodal Electronic Health Records}},
  year = {2025},
  eprint = {2507.20993},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.20993},
}
Figure