PROJECT SUMMARY/ABSTRACT Acute respiratory distress syndrome (ARDS) is a common and devastating cause of acute respiratory failure. There are 200,000 annual ARDS cases in the U.S. (2.5-5 million globally), which account for 60,000 deaths and enormous physical, cognitive, and psychosocial morbidity among survivors. Yet, despite more than 200 randomized clinical trials (RCTs), only two interventions – low-tidal-volume ventilation and prone positioning – have definitively improved outcomes using a traditional frequentist, null hypothesis, p-value-based trial design and analysis. The research team contends that assessing data in this framework may overlook informative trial data and delay or thwart the identification of promising therapies, especially when p-values fall just short of the 0.05 threshold, which has occurred in several major ARDS trials. As an alternative methodological approach to maximize the clinical insight gained from RCTs, the team will reanalyze 29 international and NHLBI-funded ARDS RCTs that enrolled more than 15,000 individuals using Bayesian causal inference and machine learning methods they have developed and validated. Most therapies they will examine are either low-cost or easily implemented practices and thus have the potential for high impact (e.g., ventilator settings, fluid management, corticosteroids, statins, beta-agonists, vitamin D). In Aim 1, instead of using statistical significance, they will quantify the probability of a beneficial treatment effect and its probable magnitude. That is, instead of using a pre-specified p-value to determine whether an intervention has at least the hypothesized mortality benefit, they will derive the probability that a given therapy is associated with clinically relevant absolute mortality reductions of at least 2%, 4%, and 6%. They will examine each intervention with noninformative Bayesian ‘priors’ and then with standardized and meta-analysis-derived priors to reduce subjectivity and interrogate clinical efficacy across the spectrum of harm and benefit possibilities. In Aim 2, they will use Bayesian Additive Regression Trees (BART) formulations they developed to understand which ARDS patient types are most likely to benefit from, or be harmed by, a therapy, i.e., so-called ‘heterogeneity of treatment effect’ (HTE). Unlike prior HTE research in ARDS, their approach does not focus on one-by-one, binary splits of characteristics but rather can identify complex, multivariable, nonlinear treatment effect modification. Aim 2a will focus on mortality and adverse events. Aim 2b will apply a novel BART variation to identify HTE in outcomes such as ventilator duration or hospital stay whose observation is truncated by death. By estimating causal effects on these outcomes among always-survivors, their new method avoids biases associated with prior approaches, enabling accurate identification of clinically meaningful subgroups. Aim 3 focuses on developing and disseminating free, cloud-based software to support future ARDS trials. This work promises to improve the value of the knowledge gained from past and future ARDS RCTs by identifying truly beneficial treatments and informing how these therapies can be individually tailored for this high-mortality, high-morbidity syndrome.

PROJECT NARRATIVE Acute respiratory distress syndrome (ARDS) is a common condition of severe respiratory failure precipitated by a variety of insults (e.g., Covid-19 and other respiratory viruses, sepsis, and pneumonia), with hospital mortality of 30–40% and significant morbidity among survivors. ARDS randomized clinical trials (RCTs) have been hampered by (1) analyses that yield a strict binary conclusion concerning treatment efficacy, (2) inadequate statistical methods to assess heterogeneity of treatment effect among varying patient types, and (3) death- truncated non-mortality outcomes such as length of stay. To offer more intuitive probabilistic interpretations of an intervention’s efficacy and thus maximally leverage the information learned from a trial, the researchers will apply a suite of Bayesian causal inference and machine learning methods to 29 international and NIH/NHLBI- funded multicenter ARDS RCTs to (1) determine the probability that the tested interventions produce benefit or harm on the absolute and relative scale for both mortality and non-mortality clinical outcomes, (2) quantify the degree of confidence in these conclusions, and (3) identify clinically meaningful subgroups most likely to benefit.