Use of Wearable Technology and Deep Learning to Improve the Diagnosis of Brugada Syndrome

Shun Liao; Mahmoud Bokhari; Praloy Chakraborty; Adrian Suszko; Gavin Jones; Danna Spears; Michael Gollob; Zhaolei Zhang; Bo Wang; Vijay S Chauhan

doi:10.1016/j.jacep.2022.05.003

Use of Wearable Technology and Deep Learning to Improve the Diagnosis of Brugada Syndrome

JACC Clin Electrophysiol. 2022 Aug;8(8):1010-1020. doi: 10.1016/j.jacep.2022.05.003. Epub 2022 Jun 29.

Authors

Affiliations

¹ Peter Munk Cardiac Center, Division of Cardiology, Toronto General Hospital, University Health Network, Toronto, Canada.
² Department of Computer Sciences, University of Toronto, Toronto, Canada.
³ Peter Munk Cardiac Center, Division of Cardiology, Toronto General Hospital, University Health Network, Toronto, Canada. Electronic address: bo.wang@uhnresearch.ca.
⁴ Peter Munk Cardiac Center, Division of Cardiology, Toronto General Hospital, University Health Network, Toronto, Canada. Electronic address: vijay.chauhan@uhn.ca.

PMID: 35981788
DOI: 10.1016/j.jacep.2022.05.003

Abstract

Background: The diagnosis of Brugada syndrome by 12-lead electrocardiography (ECG) is challenging because the diagnostic type 1 pattern is often transient.

Objectives: This study sought to improve Brugada syndrome diagnosis by using deep learning (DL) to continuously monitor for Brugada type 1 in 24-hour ambulatory 12-lead ECGs (Holters).

Methods: A convolutional neural network was trained to classify Brugada type 1. The training cohort consisted of 10-second standard/high precordial leads from 12-lead ECGs (n = 1,190) and 12-lead Holters (n = 380) of patients with definite and suspected Brugada syndrome. The performance of the trained model was evaluated in 2 testing cohorts of 10-second standard/high precordial leads from 12-lead ECGs (n = 474) and 12-lead Holters (n = 716).

Results: DL achieved a receiver-operating characteristic area under the curve of 0.976 (95% CI: 0.973-0.979) in classifying Brugada type 1 from 12-lead ECGs and 0.975 (95% CI: 0.966-0.983) from 12-lead Holters. Compared with cardiologist reclassification of Brugada type 1, DL had similar performance and produced robust results in experiments evaluating scalability and explainability. When DL was applied to consecutive 10-second, clean ECGs from 24-hour 12-lead Holters, spontaneous Brugada type 1 was detected in 48% of patients with procainamide-induced Brugada syndrome and in 33% with suspected Brugada syndrome. DL detected no Brugada type 1 in healthy control patients.

Conclusions: This novel DL model achieved cardiologist-level accuracy in classifying Brugada type 1. Applying DL to 24-hour 12-lead Holters significantly improved the detection of Brugada type 1 in patients with procainamide-induced and suspected Brugada syndrome. DL analysis of 12-lead Holters may provide a robust, automated screening tool before procainamide challenge to aid in the diagnosis of Brugada syndrome.

Keywords: Brugada syndrome; Holter ECG; diagnosis; machine learning.

MeSH terms

Brugada Syndrome* / diagnosis
Deep Learning*
Electrocardiography / methods
Humans
Procainamide
Wearable Electronic Devices*

Substances

Procainamide