COVID-19-related ARDS Supported with Extracorporeal Membrane Oxygenation: Using Machine Learning Models to Improve Care

ABSTRACT

Background: Revised guidelines clarify indications for extracorporeal membrane oxygenation (ECMO) support in patients with COVID-19-related acute respiratory distress syndrome (ARDS). Commercially available ECMO analytics software records granular perfusion data continuously throughout the run. To date, electronic-medical record (EMR) clinical data has not been integrated with ECMO perfusion data and analyzed with machine learning-based algorithms to improve patient care. Methods: Retrospective chart review was performed on all SARS-CoV2 positive patients cannulated to veno-venous ECMO at an urban highvolume regional referral center from March 1st, 2020, through December 31st, 2021. Categorical data including patient demographics, clinical outcomes, and laboratory data (complete blood count, basic metabolic panel, arterial blood gas, lactate, anticoagulation assays) and vital signs (pulse, arterial line blood pressure, oxygen saturation) were collected for the entirety of the ECMO run. Time-series perfusion data (arterial flow normalized to body surface area (BSA), sweep gas, delta pressures normalized to arterial flow) were captured every 60-120 seconds. We constructed a predictive long-short term memory (LSTM) predictive model that integrated clinical and time-series data using an extended machine learning (ML) framework with neural network. Primary outcome was successful ECMO decannulation. Data were truncated to discrete and relative timepoints (7, 14, 21 days, or percent of the run). Receiver operating characteristic (ROC) curves show the model's diagnostic accuracy. Results: 42 patients were included in the analysis (30 male, 12 female). Mean age was 43.9 (SD=11.5) years old, and mean duration of ECMO run was 36.2 (SD=30.1) days. 24 patients were successfully decannulated and 4 are currently supported on ECMO. When provided the complete data, the LSTM model showed an area under the ROC curve >0.95, demonstrating strong diagnostic accuracy in predicting successful ECMO decannulation (Figure 1A). When data were truncated to only the first two weeks of the ECMO run, the area under the ROC curve was 0.93 (Fig. 1B). Patterns of arterial flow normalized to BSA and sweep gas normalized to flow also appear different in patients with divergent clinical outcomes (Fig 2). Conclusion: Characterizing key determinants of ECMO support may offer intensive care unit healthcare teams potentially lifesaving information in real-time. Our machine-learning model successfully integrates clinical and perfusion data from the mind's eye of a clinician managing the care of a patient supported with ECMO. We have identified critical variables with the most meaningful impact on the mechanics of ECMO support. Our model may also help predict patient outcomes into and offer clinicians opportunities for interventions to improve care. (Figure Presented).