Human oversight ECG monitors outperform AI for arrhythmia detection
Click Here to Manage Email Alerts
Long-term continuous ECG monitoring with human oversight outperformed an artificial intelligence-based mobile cardiac telemetry monitor for detection of arrhythmias, researchers reported at Heart Rhythm 2021.
Detection and preliminary interpretation of results from outpatient cardiac monitors is performed by the device’s manufacturer, “so we sought to compare the accuracy of two real-world monitors that we use in our clinic,” Mark E. Willcox, MD, electrophysiologist at Alaska Heart & Vascular Institute, Anchorage, said during a press conference.
The monitors were a mobile cardiac telemetry system (BioGuardian MCT/CEM, Preventice Solutions) and a long-term continuous ECG monitor (Carnation Ambulatory Monitor, Bardy Diagnostics).
“The fundamental differences between these monitors boil down to how they are analyzed after the full disclosure data arrive at an independent data testing facility, which processes the raw data and turns it into strips that are interpretable,” Willcox said. “Preventice’s MCT/CEM does this by using a computerized algorithm that scans the full disclosure. If it detects arrhythmias, those arrhythmias are then interpreted by a human being and put into the final report. Long-term ECG has a human reviewing the full disclosure data, pulling out what is relevant and important. There is an algorithm that then goes behind that human and increases sensitivity. In the final report, they are presented differently.”
Fifty patients who had a 30-day mobile cardiac telemetry system ordered were asked to wear the long-term continuous ECG monitor simultaneously. Two electrophysiologists interpreted and adjudicated all reports and compared all adjudicated arrhythmias. Neither company’s readers were aware of the trial, Willcox said.
The results included data from 46 patients (mean age, 58 years; 33% men) who wore both monitors for a mean of 10.3 days, and in 43% of them, the mobile cardiac telemetry system defaulted to cardiac event monitoring, Willcox said during the press conference.
There were 46 adjudicated arrhythmia events during simultaneous wear, of which 24% were diagnosed by the mobile cardiac telemetry system and 50% were diagnosed by the long-term continuous ECG monitor (P = .018), according to the researchers.
The number of arrhythmia episodes detected during simultaneous wear was 19 for the mobile cardiac telemetry system and 61 for the long-term continuous ECG monitor (P < .001), Willcox said.
The mobile cardiac telemetry system detected one arrhythmia missed by the long-term continuous ECG monitor, whereas the long-term continuous ECG monitor detected 46 arrhythmias missed by the mobile cardiac telemetry system, he said.
“Not all monitors are created equal,” Willcox said during the press conference. “We all know from reading ECGs that computers aren’t quite as good as human beings, yet we rely on them a lot in the outpatient world. We are not trying to say that one monitor is clinically better than the other, because we are not comparing the clinical utility, only the diagnostic accuracy.”
The arrhythmias detected in the study were atrial fibrillation of at least 10 seconds, atrial flutter of at least 10 seconds, atrial tachycardia of at least 20 beats, ventricular tachycardia of at least three beats of 100 bpm of greater, second-degree atrioventricular block, complete heart block, sinus bradycardia of less than 30 bpm for at least 30 seconds, 2:1 sinus node exit block, atrioventricular nodal reentrant tachycardia and atrioventricular reentrant tachycardia.
Perspective
Back to Top
The study showed us that humans can still outperform machines when it comes to identifying rhythm disturbances. The authors demonstrated that continuous ECG monitoring that is overread by humans performed significantly better than algorithm-dependent mobile cardiac telemetry. The authors reiterate the point that diagnostic accuracy is often overlooked in the interests of ease of use and clarity of report while using the algorithm driven monitoring systems.
I think this should make us vigilant and examine the mobile cardiac telemetry reports carefully. We should always make sure to clinically correlate the symptoms of a patient with the ECG strips, as sometimes ECGs read by the algorithms may be false negatives.
The study needs to be repeated in larger population cohorts, with more monitors and in a wide variety of patients, with variable baseline ECG features. Primarily to make it more generalizable.
The future is nevertheless going to be AI- or algorithm-derived adjudications of rhythm disturbances. We will get there, but it will be iterative. Competition between humans and machines is healthy, but the future is about collaboration.
Collapse
Willcox ME, et al. LBCT02-01. Presented at: Heart Rhythm 2021; July 28-31, 2021 (hybrid meeting).
Read more about
Click Here to Manage Email Alerts