Individualized Prediction of Treatment Effects Using Data from Both Embedded Clinical Trials and Electronic Health Records

点击翻译按钮获取中文摘要

Abstract More than 790,000 patients undergo mechanical ventilation for acute respiratory failure (ARF) in the United States each year at a cost of $27 billion. The in-hospital mortality for these patients is nearly 35%, and for patients with critical illness, such as acute respiratory distress syndrome (ARDS), mortality can approach 50%. In some patients, guideline-appropriate care with lung-protective ventilation or prone positioning will save lives, yet in many others, an individualized treatment is elusive. There is a need for advances in leveraging opportunities in data science to improve outcomes from respiratory failure. The primary method for generating new evidence is the randomized clinical trial (RCT). Yet they are often costly, take many years, and can be slow to accelerate learning and implementation at the bedside. In addition, RCTs usually enroll a moderate number of patients at high cost (100 to 1000s) and measure a limited range of covariates (10 to 100s). Thus, they do not lead to prediction of highly individualized treatment effects, as called for by the NHLBI Working Group on Research Priorities. In contrast, real-world evidence from electronic health records (EHRs) includes many patients (often millions) and covariates (often 1000s). They are inherently generalizable, less costly, and less timely to acquire than conducting RCTs. However, the estimation of treatment effects from EHR data is often biased due to confounding, which occurs when a treatment and its effect(s) are both causally influenced by one or more events. This project uses two Specific Aims to solve these challenges. Aim 1 proposes to develop and evaluate a new method for making individualized predictions of treatment effects using data from RCTs and EHRs. It uses “embedded” RCTs in which the clinical trial occurs within the context of usual care of a health system. The embedded RCT data are applied to control for confounding when using EHR data to predict treatment effects. Aim 2 will apply these methods to two embedded RCTs at UPMC that are studying treatments that may help prevent ARF. The OPTIMISE C-19 trial is studying monoclonal antibody therapy for non-hospitalized patients with SARS-CoV-2 infection. The PeriOp trial will be studying perioperative interventions to improve post-operative outcomes after major surgery. The hypothesis to be investigated is that the proposed new methods will predict the effects of treatment on acute respiratory failure and other outcomes more accurately than will using the clinical trial or the EHR data alone. Such results would provide support that these methods yield individualized predictions of treatment effects that can inform clinical care to help prevent ARF.