Project Summary
In-hospital cardiac arrests (IHCA) occur in over 350,000 adults in the United States each year, or one IHCA
every two minutes. Despite advances in management, patient survival rates with favorable neurologic function
remain low. Ventricular tachycardia (VT) is the most common initial shockable cardiac rhythm identified at the
onset of an IHCA and the most important determinant of survival is defibrillation in <3 minutes. While,
continuous electrocardiographic (ECG) monitoring is the non-invasive gold standard method used to identify
VT, false alarms are extremely common. False VT not only impedes immediate identification of true VT and
lifesaving defibrillation, but contributes to alarm fatigue in clinicians. A widely held belief is that poor skin
electrode contact and/or default monitor settings cause false VT. As a result, clinical scientists have tried a
variety of interventions (e.g., customize alarm settings; daily skin electrode changes; disposable lead wires;
education) to decrease the number of false alarms. However, we found that fewer than 10% of false arrhythmia
alarms were due to poor signal quality (i.e., unanalyzable due to excessive noise, baseline wander, leads off).
Rather, our prior research shows that the vast majority of false alarms are due to poorly designed VT
algorithms. At the heart of the problem, are outdated databases used by monitoring manufacturers to develop
and test VT algorithms for use in bedside ECG monitors. Therefore, improvements to VT algorithms for use in
21st century bedside monitors has been stalled for decades. Recently, our group completed a multi-tiered,
multi-expert, ground truth, manual annotation, with three-person ascertainment of VT events testing a new VT
algorithm created by our group. The UCSF VT database represents the single largest human annotated
database in an intensive care unit (ICU) cohort in existence. We now aim to move these extensive efforts
forward to augment our original VT algorithm using a data driven artificial intelligence approach and increase
the generalizability of our VT algorithm by including step-down/telemetry unit patients. The specific aims are:
Aim 1. Leverage our human annotated VT database and machine learning (ML) approaches to identify novel
ECG features to create an “optimized” VT algorithm to predict VT associated and IHCA. Aim 2. Compare the
following VT algorithms: (1) v1 (signal processing); (2) ML/AI (Aim 1), and the hospital-based ECG monitors
(i.e., ICU and step-down/telemetry unit) using prospective data in 5,000 ICU (50% of total) and step-
down/telemetry patients (50% of total). Designing and testing clinically relevant VT algorithms that both
improve identification of true VT and forecast associated IHCA has important implications for reducing
preventable morbidity and mortality, reducing alarm fatigue, improving patient safety, enhancing nursing care
and ECG monitoring systems.
