Harnessing Machine Learning Models to Predict Outcomes in Patients Supported with Extracorporeal Membrane Oxygenation
Video
Team Information
Team Members
Joshua Fuller, Undergraduate Student in Biomedical Engineering, Columbia Engineering
Alexey Abramov, MD, Department of Surgery, Columbia University Irving Medical Center
Dana Mullin, Clinical Perfusion, New York Presbyterian Hospital
Philippe Lemaitre, MD, Department of Surgery, Columbia University Irving Medical Center
Faculty Advisor: Elham Azizi, Assistant Professor of Biomedical Engineering, Columbia Engineering; and Florence Irving Assistant Professor of Cancer Data Research, Irving Institute for Cancer Dynamics
Abstract
Extracorporeal membrane oxygenation (ECMO) machines are miniaturized heart lung bypass machines that are increasingly used to support patients with respiratory failure or refractory cardiogenic shock. Modern ECMO devices with analytics software allow the recording of multiple machine variables during the support run (perfusion data). On the patient side, modern electronic medical records also allow capture of granular data during a hospital stay (patient metadata), such as clinical variables and laboratory values. The goal of this project was to analyze ECMO perfusion data and patient metadata using Long-Short-Term Memory (LSTM) neural networks to discriminate between successful and unsuccessful decannulation. The perfusion and laboratory data from thirty six COVID-19 ECMO patients was collected, and used to create two LSTM models. The first was only given the twelve perfusion variables, whereas the second was given the twelve perfusion variables and seventeen lab variables. The full data set was used to measure the feasibility of each model. Once that was established, each model was trained on limited time periods: first week on ECMO, first two weeks on ECMO, and first three weeks on ECMO. To measure the performance of each of the models, the area under the receiver operator curve (AUC) was calculated. Both models showed AUC above 0.94 when trained on the full data set, demonstrating that an LSTM can discriminate between unsuccessful and successful decannulation. When the data was truncated, the model’s predictive power weakens. However, the LSTM with only laboratory values is still more accurate than random chance, even with just one week of data. When laboratory values are added, accuracy increases significantly. Two weeks of perfusion and laboratory data has more diagnostic power than three weeks of perfusion data alone. This signals that the LSTM can incorporate clinical information into its classification, further demonstrating its usefulness as a clinically predictive tool.
Team Lead Contact
Joshua Fuller: jsf2168@columbia.edu
