Emulated target trials and phenotyping in patients with acute respiratory distress syndrome

The National Heart, Lung, and Blood Institute (NHLBI) awarded Massachusetts Institute of Technology $1.48 million under the Cardiovascular Diseases Research program (CFDA 93.837) to develop advanced analytical and machine learning methodologies for evaluating treatment strategies in acute respiratory distress syndrome (ARDS). The grant, awarded September 23, 2025, with completion targeted for August 31, 2027, will deliver research products addressing critical gaps in ARDS phenotyping and personalized treatment optimization. The project will employ target trial emulation techniques and G-methods for confounding adjustment to evaluate dynamic treatment regimes—sequential treatment decision rules based on evolving patient clinical data—using three large observational datasets: the Medical Information Mart for Intensive Care (MIMIC) IV Database, the eICU Collaborative Research Database, and the Dutch AMSTE dataset.

The research will generate machine learning-based dynamic phenotyping methods for ARDS patient prognostication and personalized treatment recommendations, focusing on commonly applied interventions including neuromuscular blockade, steroids, driving pressure, and mechanical power that currently lack high-quality prospective evidence. By leveraging existing real-world clinical data rather than conducting infeasible prospective randomized trials of all potential treatment strategies, the project will deliver methodological frameworks and evidence to inform clinical decision-making regarding which ARDS interventions benefit specific patient populations and optimal treatment timing.

Abstract

Acute respiratory distress syndrome (ARDS) is a severe form of lung injury with significant public health implications due to severe morbidity and mortality. The need to utilize existing data to inform prospective research and clinical decision making was emphasized during the covid pandemic, when ARDS became a leading cause of death, and clinicians were forced to operate outside of existing evidence. Many commonly applied interventions including use of neuromuscular blockade (NMB), steroids, driving pressure and mechanical power are used despite negative data, without high quality prospective studies or with equivocal evidence and with the potential for both benefit or harm. Additionally, phenotypes of ARDS may have different prognosis and response to treatment, but thus far have not been well differentiated using routinely available dynamic clinical data, nor have they been incorporated into prospective trials. 'dynamic treatment regimes' (DTRS) are rules for making treatment decisions sequentially at multiple time-points based on a patient's evolving history. Most relevant treatment strategies for ARDS are DTRS. DTRS may be evaluated in randomized trials, however IT is infeasible to conduct randomized trials testing all DTRS of interest. This grant proposes 'target trial emulation' from observational data using 'G-methods' for confounding adjustment to address multiple gaps in our knowledge about ARDS care. We will address important methodological gaps in ARDS phenotyping and develop advanced machine learning (ML) methods for dynamic phenotyping for prognostication and personalized DTRS to determine for whom and when specific ARDS treatments are beneficial. The investigation will utilize three large datasets-including the medical information mart for intensive care (MIMIC) IV database, the eicu collaborative research database, and the dutch amsterdamumcdb database-representing a wide geographic and demographic spectrum, and the ability to assess stability of findings across geography and centers. To address these knowledge gaps regarding use of NMB, steroids, driving pressure and mechanical power, as well as identify phenotypes of patients most responsive to treatment, we propose two overarching aims for this grant. Specific aim 1) using target trial emulations and G-methods, we will estimate clinical outcomes that would result under a range of treatment strategies for NMB, steroids as well as driving pressure and mechanical power thresholds. Specific aim 2) we will develop machine learning methods to derive dynamic markers for ARDS phenotyping and formulate personalized DTRS for ARDS treatment. The project represents a collaborative effort between experts in critical care medicine (with a specialty in mechanical ventilation), machine learning, and causal inference. Our results will address important gaps in clinical knowledge about treatment of ARDS and inform the design of future randomized trials. Our study designs, code, and constructed cohorts will also provide valuable templates for other researchers to use in future observational studies, which we foresee will broadly improve the quality of evidence from observational data in critical care.