Decisions in hospitals affect patient outcomes and hospital finances alike. Examples include when to discharge a patient, when to initiate or end invasive ventilation, and whom to prioritize for monitoring and treatment. Such decisions are made under uncertainty and pressure, often relying on experience and a limited set of patient characteristics, while hospital information systems – especially in intensive care medicine – record thousands of features per patient that remain unused. The consequences of suboptimal decisions are substantial: premature discharges, delayed admissions, and false prioritization worsen patient outcomes, increase costs, and lower productivity. This issue is especially serious given that hospitals face constrained staff resources, rising cost pressure, and increasing demand due to demographic change, while simultaneously being expected to deliver high-quality care.
We hypothesize that the precision medicine approach of tailoring treatments to individual patients, rather than relying on averages, could prove effective for managing decision-making under uncertainty. Our idea is to employ causal machine learning (CML) methods – particularly causal forests, which combine econometrics with machine learning – to this end. Recent methodological advances in CML enable the estimation of individualized treatment effects from observational data, a prerequisite for personalized decisions. By evaluating past decisions in retrospective data, CML can help identify which decision is optimal for which patient.
Despite these advances, empirical applications of CML to real-world hospital decision-making remain rare. We address this gap. Building on our reference study of intensive care unit (ICU) discharge decisions, we will develop and test an empirical framework for applying CML to hospital operations and decision-making. We will focus on two types of high-impact decisions: (1) treatment choices within the ICU, such as initiation or termination of mechanical ventilation, and (2) patient-flow decisions to and from the ICU, such as primary ICU admission and ICU discharge timing.
Using retrospective data from tens of thousands of patients across four partner hospitals, described by several thousand features, we will train causal forests to estimate individualized average treatment effects and policy trees to estimate group average treatment effects. The performance of learned policies will be evaluated through simulation experiments on unseen data, comparing their utility in terms of avoided readmissions, reduced complications, saved ICU capacity, and improved operating profit per case.
A multicenter setting allows us to test whether hospital-specific models, potentially capturing organizational behaviors, perform better than pooled models that leverage larger sample sizes. With this setting, we aim not only to contribute to the literature but also to inform future practical implementation: should CML models for decision support be trained centrally, or should each hospital train its own?
Our project is use-inspired. Practitioners express a clear need for data-driven tools that support personalized decisions under uncertainty. Accordingly, our team comprises senior clinicians and experienced medical researchers alongside experts in applied econometrics and data science, healthcare management science, and hospital management.
If successful, our research could demonstrate that medical and economic outcomes can be improved simultaneously through personalized decision-making, marking a paradigm shift for both medicine and operations management. We aspire that our project will motivate other researchers to adapt our framework to other data-rich industries where uncertainty and limited resources prevail.
Marco Samorí, Martina Vecchione, Andrés Gallardo (as of March 2027), Dr. Justus Vogel (Principal Investigator)
Health Ostschweiz (HOCH) – Cantonal Hospital of St. Gallen: Prof. Dr. med. Urs Pietsch (Co-Investigator) and Prof. Dr. med. Miodrag Filipovic
University Hospital Zurich: Prof. Dr. Reto Schüpbach
Bern University Hospital: Prof. Dr. Carmen Pfortmüller
University Hospitals of Geneva: Dr. David Legouis
Swiss National Science Foundation (see the official SNSF publication here)
48 months (planned start / timeline: October 2026 – September 2030)