TREAT ECARDS: Translating Evidence into Action: Electronic Clinical Decision Support in Acute Respiratory Distress Syndrome

Project Final Report (PDF, 1.6 MB) Disclaimer

A clinical decision support system that uses machine learning combined with clinician perspectives to identify and manage patients with acute respiratory distress syndrome is feasible and outperforms clinician recognition.

Project Details -
Completed

Acute respiratory failure (ARF) is the most common acute organ dysfunction in US hospitals, with acute respiratory distress syndrome (ARDS) being the most severe form with a mortality rate of 40 percent. Historically, only one third of those with ARDS are recognized, contributing to the under-utilization of evidence-based practices (EBPs) known to improve outcomes.

This research developed and evaluated the effectiveness of a clinical decision support system called TRanslate Evidence into AcTion (TREAT) that identifies and supports management of ARDS. TREAT is composed of the ARDS Sniffer 2.0 that identifies patients with ARDS and a clinical dashboard called ECARDS--Electronic Clinical decision support in ARDS. ECARDS tracks the severity of illness and likelihood of ARDS in real time and tracks outcomes and electronic health (EHR) record-based performance measures (EPMs) for intensive care units.

The specific aims of the research were as follows:

  • Develop and validate an automated, EHR-based tool to identify patients with ARDS.
  • Develop an evidence-based, context-appropriate system, ECARDS, for the management of ARDS and an ARDS Dashboard for audit and feedback.
  • Evaluate the effectiveness of the ARDS Sniffer 2.0 and ECARDS in a real-world clinical setting.
  • Promote the dissemination of ARDS Sniffer 2.0 and ECARDS through partner professional organizations.

A systematic review of ARDS and ARF along with interviews of clinicians helped identify six EBPs for ECARDs. To measure and track the chosen EBPs, the research team developed and validated EPMs. They then developed a deep learning model to identify patients with ARDS as part of ARDS Sniffer 2.0. To train the model, a historic dataset from a mechanically ventilated (MV) non-COVID-19 cohort was used to identify patients with higher risk of ARDS or in-hospital mortality. The model was then validated on a non-COVID-19 cohort with MV, a COVID-19 positive cohort with and without MV, and a subgroup of COVID-19 positive patients on MV. Separate cohorts were used because patients with COVID -19 requiring MV nearly all meet the criteria for ARDS and thus have far higher rates of recognition of ARDS than has historically been the case.

The ARDS Sniffer 2.0 identified ARDS with a sensitivity of 86 percent in non-COVID MV patients, 70 percent in COVID-19 patients with and without MV, and 92 percent in MV COVID-19 patients. In the MV COVID-19 cohort, the prevalence of ARDS was so high that the Sniffer alerted on 75 percent of patients, with nearly 100 percent of patients who triggered the Sniffer confirmed to have ARDS. Because of strain on resources during the COVID-19 pandemic, and the fact that most hospitalized patients with COVID-19 pneumonia were recognized to have ARDS, ECARDS was only tested in a simulated sandbox and not implemented into a clinical setting. The researchers concluded that the use of the ARDS Sniffer 2.0 and ECARDS holds promise as an approach to implement similar tools to assist in the identification and management of other under-diagnosed conditions.