Home > Cardiology > EHRA 2022 > Diagnostics and Prevention > AI model accurately discriminates between arrhythmias

AI model accurately discriminates between arrhythmias

Presented by
Dr Arunashis Sau, Imperial College London, UK
Conference
EHRA 2022
Doi
https://doi.org/10.55788/84de9f2e
An artificial intelligence (AI) model was able to successfully distinguish between patients with mus (CTI)-dependent atrial flutter and non-CTI dependent atrial tachycardia, based on patient ECGs. Since the model performed well in this proof-of-concept study, future studies will explore the diagnostic capacities of AI with regard to other arrhythmias.

It would be helpful for clinicians and patients if the mechanism of arrhythmias could be identified via ECGs with a high degree of certainty, according to Dr Arunashis Sau (Imperial College London, UK) [1]. Two main categories of atrial arrhythmias are CTI-dependent atrial flutter and non-CTI dependent atrial tachycardia. The current study aimed to train a convolutional neural network to discriminate between these 2 categories of arrhythmias. The model was compared with expert assessments, using an invasive electrophysiology study as the source of truth. Collected were 13,557 ECGs from 288 patients. The training data set consisted of 13,500 ECGs from 231 patients and the test set included 57 ECGs from 57 patients.

The model achieved an accuracy of 86%, which was a significantly higher accuracy than the electrophysiologist assessment (79%) and a numerically higher accuracy than the electrophysiologist consensus (81%). The area under the curve of the model was 0.94. Notably, experts were more likely to incorrectly diagnose an atrial flutter case as being atrial tachycardia than the AI model (34.5% vs 10.3%). According to Dr Sau, this finding could have significant implications, given that CTI-dependent atrial flutter is more amendable to catheter ablation. Furthermore, when the model and electrophysiologist consensus agreed, the prediction accuracy was 100%. “This result indicates that the use of this model with human-in-the-loop provides powerful results,” argued Dr Sau.

“We successfully trained a neural network to distinguish CTI-dependent atrial flutter from atrial tachycardia, with a performance that is at least equivalent to human expert performance,” concluded Dr Sau. “Other studies will be conducted to further analyse the use of AI in ECG-based diagnosing of patients with arrhythmias.”

  1. Sau A, et al. Classification of organised atrial arrythmias using explainable artificial intelligence. E-cardiology award session, EHRA 2022, 3–5 April, Copenhagen, Denmark.

Copyright ©2022 Medicom Medical Publishers



Posted on