Machine Learning Model Predicts the Likelihood of Achieving Return of Spontaneous Circulation During
Takayuki Sueishi, Hunter Gaudio, Luiz E. V. Silva, Tiffany S. Ko, Fuchiang Tsui, Ryan W. Morgan, Todd J. Kilbaugh
Poster # 77
In the United States alone, approximately 6000 children experience in-hospital cardiac arrest (IHCA) each year. There is a great need for a reliable method to determine the likelihood of achieving return of spontaneous circulation (ROSC) from cardiopulmonary resuscitation (CPR) and vasopressor administration alone. This information can be used to determine if an alternate/additional treatment pathway (i.e. extracorporeal membrane oxygenation) is required. The Resuscitation Science Center within the Children's Hospital of Philadelphia has a validated swine model of asphyxia-based cardiac arrest. Subjects whose data was considered in this study all underwent 7 minutes of asphyxia followed by the induction of ventricular fibrillation and 10 minutes of hemodynamic-guided CPR and vasopressor administration. Compression depth was targeted to a systolic arterial blood pressure (SBP) of either >100mmHg or >80mmHg depending on experimental group. Similarly, vasopressors were delivered in the following cycle, starting 2 minutes into CPR: epinephrine, wait 1 minute, epinephrine, wait 1 minute, vasopressin, wait 2 minutes. After the required wait time, vasopressors were only administered if the diastolic arterial blood pressure (DBP) fell below 30mmHg. To predict the likelihood of ROSC in real-time during CPR, a machine learning model was developed to evaluate the hemodynamic response to vasopressor administration. High resolution (100Hz) continuous DBP, SBP, and end-tidal carbon dioxide (EtCO2) data was considered 15 seconds prior to and 50 seconds after vasopressor administration. This data was divided into epochs, preliminary calculated and demographic features were selected, and a logistic regression model using 10-fold cross validation was trained on a data set of 109 subjects (97 ROSC, 12 no ROSC). Initial results yielded an area under the ROC (receiver operating characteristic) curve (AUC) score of 0.829 in a model considering only DBP and SBP from the 2nd epinephrine administration. Work is ongoing to analyze feature importance and train other machine learning models.