In the Journals

AI system performs similarly to cardiologists to diagnose arrhythmias

Mark J. Day
Mark J. Day

An end-to-end deep learning approach, a form of an AI system, classified a variety of arrhythmias through single-lead ECGs with a similar high diagnostic performance to a cardiologist, according to a study published in Nature Medicine.

“For the first time, a deep-learned algorithm has been demonstrated to provide a clinical output that is meaningful with respect to the way rhythms are interpreted by cardiologists in 12 different output classes instead of just one or two,” Mark J. Day, executive vice president of research and development at iRhythm Technologies, told Cardiology Today. “That output has been shown to be validated against what experts and cardiologists can do in terms of annotation and specifically showing that it can match them and perform this level of annotating ECG.”

Awni Y. Hannun, PhD student in computer science at Stanford University, and colleagues developed a deep neural network from a training dataset of 91,232 ECG records from 53,549 patients (mean age, 69 years; 43% women). The deep neural network was able to detect 12 output rhythm classes, including 10 arrhythmias, sinus rhythm and noise. Patients wore an ambulatory ECG monitor (Zio, iRhythm) to collect ECG data for a mean of 10.6 days.

The deep neural network was validated with a test dataset of 328 ECG records from 328 patients (mean age, 70 years; 38%) women.

The deep neural network achieved an average area under the receiver operating characteristic curve of 0.97 when validated against an independent test dataset.

Deep neural network F1 scores were similar to those of the average cardiologist, with lower scores on similar classes such as ectopic atrial rhythm and ventricular tachycardia. The set-level average F1 scores for the deep neural network were higher compared with the average cardiologist (0.837 vs. 0.78). The sensitivity of the deep neural network exceeded the average cardiologist for all rhythm classes when the specificity at the average specificity level by cardiologists was fixed.

“That in no way is to suggest that this algorithm is going to supplant or replace cardiologists,” Day said in an interview. “Rather, this is very specifically a technology to help them be much more effective in the work that they do. It is not a value to have cardiologists sifting through large amounts of ECG information, but rather to have them focused on developing and delivering clinical care programs for patients. That’s exactly what these types of algorithms can do to make their work a little more efficient and, therefore, deliver care to patients more effectively and quickly.”

The next steps for this technology include extending the duration of ECG recordings the algorithm can process, Day said.

“For practical use, we needed to extend the algorithm from processing 30-second records to the 14-day ECG recordings of iRhythm’s Zio devices, during which a typical patient’s heart beat will beat up to 1.5 million times,” Day told Cardiology Today. “iRhythm has submitted this work to the FDA and recently received 510(k) clearance for the extended algorithm.” – by Darlene Dobkowski

For more information:

Mark J. Day can be reached at media@irhythmtech.com.

Disclosures: Data annotation in this study was financially supported by iRhythm Technologies. Day is an employee of iRhythm Technologies. Hannun reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.

Mark J. Day
Mark J. Day

An end-to-end deep learning approach, a form of an AI system, classified a variety of arrhythmias through single-lead ECGs with a similar high diagnostic performance to a cardiologist, according to a study published in Nature Medicine.

“For the first time, a deep-learned algorithm has been demonstrated to provide a clinical output that is meaningful with respect to the way rhythms are interpreted by cardiologists in 12 different output classes instead of just one or two,” Mark J. Day, executive vice president of research and development at iRhythm Technologies, told Cardiology Today. “That output has been shown to be validated against what experts and cardiologists can do in terms of annotation and specifically showing that it can match them and perform this level of annotating ECG.”

Awni Y. Hannun, PhD student in computer science at Stanford University, and colleagues developed a deep neural network from a training dataset of 91,232 ECG records from 53,549 patients (mean age, 69 years; 43% women). The deep neural network was able to detect 12 output rhythm classes, including 10 arrhythmias, sinus rhythm and noise. Patients wore an ambulatory ECG monitor (Zio, iRhythm) to collect ECG data for a mean of 10.6 days.

The deep neural network was validated with a test dataset of 328 ECG records from 328 patients (mean age, 70 years; 38%) women.

The deep neural network achieved an average area under the receiver operating characteristic curve of 0.97 when validated against an independent test dataset.

Deep neural network F1 scores were similar to those of the average cardiologist, with lower scores on similar classes such as ectopic atrial rhythm and ventricular tachycardia. The set-level average F1 scores for the deep neural network were higher compared with the average cardiologist (0.837 vs. 0.78). The sensitivity of the deep neural network exceeded the average cardiologist for all rhythm classes when the specificity at the average specificity level by cardiologists was fixed.

“That in no way is to suggest that this algorithm is going to supplant or replace cardiologists,” Day said in an interview. “Rather, this is very specifically a technology to help them be much more effective in the work that they do. It is not a value to have cardiologists sifting through large amounts of ECG information, but rather to have them focused on developing and delivering clinical care programs for patients. That’s exactly what these types of algorithms can do to make their work a little more efficient and, therefore, deliver care to patients more effectively and quickly.”

The next steps for this technology include extending the duration of ECG recordings the algorithm can process, Day said.

“For practical use, we needed to extend the algorithm from processing 30-second records to the 14-day ECG recordings of iRhythm’s Zio devices, during which a typical patient’s heart beat will beat up to 1.5 million times,” Day told Cardiology Today. “iRhythm has submitted this work to the FDA and recently received 510(k) clearance for the extended algorithm.” – by Darlene Dobkowski

For more information:

Mark J. Day can be reached at media@irhythmtech.com.

Disclosures: Data annotation in this study was financially supported by iRhythm Technologies. Day is an employee of iRhythm Technologies. Hannun reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.