AI-guided Prediction and Treatment of Cardiac Arrest

Sudden cardiac arrest (SCA) remains a major cause of mortality in the United States. Despite significant efforts to improve resuscitation outcomes, survival remains poor. Moreover, SCA is often the first manifestation of underlying heart disease, after which survivors typically suffer secondary chronic cardiovascular and neurological disease due to the primary insult. Resuscitation from SCA is initiated in patients suffering from life-threatening arrhythmias, generally ventricular tachycardia/fibrillation (VT/VF) or pulseless electrical activity (PEA), that if successful, is followed by a return of spontaneous circulation (ROSC). This, however, is commonly followed by repeated rearrest due to either VT/VF, PEA, or asystole. Importantly, the mechanisms of VT/VF and PEA vary significantly, and protocolized, non-directed treatments can worsen outcomes. Currently, treatments are typically initiated after rearrest occurs and without regard to an individual patient's unique personal or arrest characteristics, factors which are known predictors of cardiac arrest outcomes. Recent studies have shown that a quicker and more targeted treatment response (i.e., shorter time to treat with rearrest type-specific treatment) for cardiac arrest can significantly improve survival and downstream chronic sequelae after SCA. Therefore, the ability to predict cardiac arrest or rearrest occurrence and cause (i.e. VT/VF and PEA) could guide early personalized therapy and ultimately improve resuscitation outcomes, which is the long-term translational goal of this effort. Additionally, early intervention could be critical in rural communities, where transport times to definitive hospital care are long, and resources are limited.

In ongoing work the investigators have recently observed in emergency medical services (EMS) patients that adding clinical parameters (e.g., age, primary arrest type) to features derived from the ECG T-wave in a machine learning (ML) model improves prediction accuracy of VT/VF and PEA (>90%). The investigators propose a feasibility study of ML-guided prediction of rearrest. The overall hypothesis is that an ECG biomarker, combined with clinical features, can predict rearrest and its cause (VT/VF or PEA), which will significantly improve time to treatment and cardiac arrest acute and chronic outcomes. The investigators further hypothesize that such technology can be readily adopted and successfully implemented by emergency responders. The investigators plan two interrelated clinical trials, one an observational trial in simulated cardiac arrest, and a second observational trial in EMS patients. The following specific aims will test these hypotheses.

Aim 1. Determine end-user performance and satisfaction with a fully automated ML-guided rearrest prediction device using simulated cardiac arrest scenarios. Emergency providers, in simulated real-time resuscitation scenarios using real rearrest ECG recordings will be studied to identify barriers and facilitators to technology implementation and adoption, as well as provider satisfaction and the perceived value of the intervention. This will be achieved by determining provider performance metrics (e.g., initiating ML-guided therapies) and using qualitative tools to assess end-user engagement, trust and perceived utility. These studies will provide foundational data for the implementation and scalability of the technology for both a clinical trial and guidance for future FDA and regulatory approval.

Aim 2. Test the accuracy of a multiclass ML model for real-time prediction of rearrest occurrence and its type (VT/VF vs. PEA) and impact on time to treatment, in an observational clinical trial of cardiac arrest patients. Leveraging an ongoing collaboration with Cleveland and community EMS, the investigators will enroll EMS cardiac arrest patients and observe for the occurrence of rearrest in a proof-of-concept observational validation study. The accuracy of the fully automated ML model to predict rearrest and its type will be determined, but treatment guidance will not be tested at this time. Providers will be blinded to the ML output, and ML will not direct treatment. Offline, prediction accuracy and estimated time to ML-guided treatment decision will be compared to actual rearrest type and time of treatment. These results will provide preliminary safety and accuracy results, sample size estimates, and recruitment and informed consent processes for a future randomized controlled clinical trial to test ML-guided rearrest treatments.