Development of Artificial Intelligence-Enabled Prediction for Optimal Timing of ECMO Circuit Change
Summary
Extracorporeal Membrane Oxygenation (ECMO) is used to support patients at the most extreme of critical illness. As with any extracorporeal device, including renal support therapies such as hemodialysis or continuous renal replacement therapy (CRRT), ventricular assist devices (VAD), and the like, ECMO requires anticoagulation to be administered so that when patients’ blood touches the varying different polymer surfaces of the tubing and membrane oxygenator, it does not immediately clot. The ECMO team closely monitors patients’ levels of anticoagulants, hematologic, and coagulation values to keep patients’ blood at just the right level so as to prevent clots, but also to prevent bleeding from the blood thinners.
Despite this expert level of care and attention to detail, it is inevitable that every ECMO circuit and membrane oxygenator will develop some level of fibrin stranding and eventually micro- and macro-thrombi. These clots lead to consumption of important coagulation factors, ultimately leading to an entity known as disseminated intravascular coagulation or “circuit DIC”. When the patient shows signs of circuit DIC, e.g. decreased fibrinogen level, increased INR/PT level, decrease platelet counts, and begins requiring more blood product transfusions to prevent further bleeding, the ECMO team is faced with the decision of when to “cut out” the existing bad ECMO circuit and hook in a new, fresh one. Other indicators help make the decision, but these are both objective lab measurements and pressure changes on the ECMO circuit, as well as subjective evaluation of the circuit by the ECMO team, looking for visible clots in the tubing and membrane oxygenator. Many data points are then integrated to make the final decision to change out the ECMO circuit, which is both a medically risky and costly procedure.
The goal of this novel project would be to use data points from the electronic medical record (EMR), including laboratory values, pressure changes from the ECMO machine, and other indicators of membrane oxygenator efficiency to create a machine-learning algorithm, where hopefully we can predict membrane failure and circuit DIC, which could provide us a window of opportunity to change the circuit prior to the development of circuit DIC, potentially saving the patient blood transfusions and preventing this risky period of bleeding from coagulation consumption.
Development of this AI-driven model will serve as a tool to supplement ECMO practitioners’ and team members’ clinical judgment as to the ideal timing of an ECMO circuit change, which will save patients from both unnecessary procedures as well as the time put at-risk utilizing less efficient, suboptimal support machinery. It is not intended to replace expert ECMO practitioners’ clinical decision making. Institution of this AI-driven model, or supervised learning algorithm, may decrease blood product utilization, decrease the number of ECMO circuit changes, decrease cost, decrease length of days supported with ECMO, decrease length of ICU stay, and decrease hospital length of stay. It is anticipated that collectively these may improve morbidity and mortality of patients supported with ECMO.
Keywords:
- artificial intelligence
- machine learning
- clinical informatics
- electronic health record (EHR)
Researchers:
- Matthew Malone (Author)
- amir mian
- Mandana Rezaeiahari