From Observational Data to Clinical Recommendations: A Causal Framework for Estimating Patient-level Treatment Effects and Learning Policies

Rom Gutman, Shimon Sheiba, Omer Noy Klein, Naama Dekel Bird, Amit Gruber, Doron Aronson, Oren Caspi, Uri ShalitarXiv 2025 · DOI · arXiv

Abstract

Bridging the gap between theoretical causal inference and clinical deployment, this paper introduces the Target Recommendation System (TRS). TRS is an end-to-end framework designed to safely learn personalized treatment policies from observational Electronic Medical Record (EMR) data. By integrating uncertainty-based deferral mechanisms and semi-synthetic simulations, the framework provides a rigorous blueprint for building clinical AI that improves patient outcomes over current medical practice while maintaining safety through a "defer to human" approach.

Introduction

Electronic Medical Records (EMRs) provide a vast repository of data for personalized medicine. However, translating this data into Individualized Treatment Rules (ITRs) is fraught with challenges. Standard machine learning models often capture correlations rather than causal relationships. Recommending a treatment requires estimating counterfactuals—predicting what would happen if a patient received a different treatment than the one recorded.

Observational data is inherently biased by hidden confounders that influence both treatment choice and outcomes. While research into Conditional Average Treatment Effects (CATE) has grown, a significant gap remains in creating safe, deployable clinical pipelines. The TRS framework addresses this by providing a step-by-step methodology to ensure causal identifiability, handle statistical uncertainty, and validate policies before clinical implementation.

The TRS Framework

The authors propose a three-phase framework inspired by the "Target Trial" paradigm, which emulates a randomized controlled trial using observational data. This structured approach ensures that the resulting recommendations are both statistically sound and clinically relevant.

The Target Recommendation System (TRS) framework workflow.

Phase 1: Identification

This phase involves defining a precise clinical decision point and ensuring that causal assumptions—specifically ignorability (no unmeasured confounders) and common support (patients having a non-zero probability of receiving any treatment option)—are met.

Phase 2: Estimation and Deferral

  • Propensity Score Modeling: Estimating the probability of treatment given patient features to map the "common support" region.
  • Semi-Synthetic Simulation: A critical debugging step where outcomes are simulated using real patient covariates. This validates whether the data and models can recover a known optimal policy before trusting unobservable counterfactual estimates.
  • CATE Modeling: Utilizing meta-learners (e.g., T-learners with XGBoost or Ridge regression) to estimate individualized treatment effects.
  • Deferral Set Estimation: A core safety feature where the system defers to human clinicians when a patient falls outside the common support region or when causal uncertainty is too high.

Phase 3: Validation

The final phase uses Doubly Robust (DR) and Inverse Propensity Weighting (IPW) estimators on held-out data to calculate the "policy value"—the expected average outcome of the new system compared to current clinical practice.

Clinical Case Study: ADHF and AKI

The framework was tested on a high-stakes dilemma: managing diuretics for patients with Acute Decompensated Heart Failure (ADHF) who develop Acute Kidney Injury (AKI). This scenario is a clinical "tug-of-war" between saving the kidneys (decreasing diuretics) and saving the heart (increasing diuretics).

  • Data: 2,157 hospital stays from a large university medical center (2004–2015).
  • Treatment: Decrease vs. Increase/Maintain diuretics dosage.
  • Primary Outcome: Return to Baseline (RTB) of creatinine levels (kidney recovery).

Key Results

  • Performance Gains: Current clinical practice achieved a 22.1% RTB rate. The TRS-learned policies using Ridge and XGBoost T-learners achieved 45.8% and 40.9% RTB, respectively—nearly doubling the success rate.
  • Safety and Balance: The XGBoost-based policy improved kidney function while simultaneously achieving lower 30-day re-hospitalization rates, demonstrating a successful balance between competing clinical needs.
  • The Power of Deferral: By allowing the model to say "I don't know" in high-uncertainty cases, the system significantly improved the safety and average outcomes for the patients it did choose to treat.

Discussion and Limitations

A central insight of this work is that a recommendation system does not need to be perfect for every patient to be useful. By leveraging the "Wisdom of the Crowd" (using propensity scores as a baseline) and incorporating a deferral mechanism, researchers can build systems that augment clinical expertise rather than replace it.

However, limitations remain. The current framework focuses on binary treatment decisions and does not yet account for SUTVA violations or sequential, dynamic treatment regimes. Future work will focus on multi-arm treatment setups, continuous dosages, and prospective validation in real-world hospital settings.

Summary updated

Abstract

We propose a framework for building patient-specific treatment recommendation models, building on the large recent literature on learning patient-level causal models and inspired by the target trial paradigm of Hernan and Robins. We focus on safety and validity, including the crucial issue of causal identification when using observational data. We do not provide a specific model, but rather a way to integrate existing methods and know-how into a practical pipeline. We further provide a real world use-case of treatment optimization for patients with heart failure who develop acute kidney injury during hospitalization. The results suggest our pipeline can improve patient outcomes over the current treatment regime.
BibTeX
@misc{gutman2025from,
  author = {Rom Gutman and Shimon Sheiba and Omer Noy Klein and Naama Dekel Bird and Amit Gruber and Doron Aronson and Oren Caspi and Uri Shalit},
  title = {{From Observational Data to Clinical Recommendations: A Causal Framework for Estimating Patient-level Treatment Effects and Learning Policies}},
  year = {2025},
  eprint = {2507.11381},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.11381},
}
Figure