A ‘needle in a haystack’ ECG signature of subtle atrial fibrillation during sinus rhythm can be accurately detected with an artificial intelligence (AI) algorithm, a US study shows.
The ‘ground-breaking’ results may have applications in patients with unexplained stroke (embolic stroke of undetermined source – ESUS), or heart failure, according to Australian experts.
In a retrospective study, researchers at the Mayo Clinic used a ‘deep learning’ AI algorithm on almost 650,000 10-second, 12-lead ECGs obtained from 181,000 patients with normal sinus rhythm to train a neural network to recognise the subtle differences associated with AF.
They divided the data into patients who were either positive or negative for atrial fibrillation and used 70% of the ECGs initially as a training dataset to train the network.
Evaluation of an internal validation dataset (10% of the patient cohort) was done to optimise the network, and a testing dataset (20% of the patient cohort) to identify the ability of the AI-enabled ECG to detect atrial fibrillation.
They found that based on a single AI-enabled ECG, the neural network had an area under the curve (AUC) of 0·87, a sensitivity of 79%, specificity of 79·5%, F1 score of 39·2% and overall accuracy of 79·4%.
Accuracy was improved to 83·3% when the AI algorithm was applied to multiple ECGs acquired from the same patient within a month with an AUC of 0·90, sensitivity of 82·3% , specificity to 83·4% and F1 score of 45·4%.
Writing in the Lancet, the researchers said the findings showed that the screening accuracy of AI with the “convolutional neural network” compared favourably with other screening tests such as BNP for heart failure (AUC 0·60–0·70), and the CHA2DS2-VASc Score for stroke risk (AUC 0·57–0·72).
They said that with training on thousands of sinus rhythm ECGs, the AI algorithm was able to detect wavelets on the ECG smaller than the readily observable P wave, that might reflect regional non-sinus electrical activity in patients with early structural changes that precede AF.
These might include myocyte hypertrophy, fibrosis, and chamber enlargement, and inter-atrial block (ie, Bayés syndrome), which has been shown to correlate with both risk of incident atrial fibrillation and stroke.
“It is possible that a neural network trained with exposure to more than 500,000 ECGs and with sufficient depth to extract and recall subtle features not routinely appreciated or formally reported by human observers might be powerful enough to identify such features,” they wrote.
“The ability to identify patients with potentially undetected atrial fibrillation using an inexpensive, non-invasive, widely available, point-of-care test has important practical implications for atrial fibrillation screening and potentially for the management of patients with prior stroke of unknown cause
A linked commentary co-authored by Dr Jeroen Hendriks of the Centre for Heart Rhythm Disorders, South Australian Health and Medical Research Institute, University of Adelaide said the Mayo Clinic findings were exciting and with further refinement the neural network could be used for primary atrial fibrillation prediction.
It would be particularly helpful in populations where currently continuous ECG monitoring by means of loop recorders is necessary such as embolic stroke of undetermined source (ESUS).
“Their findings will be of clinical importance, especially in identifying silent atrial fibrillation, and might have important implications for secondary prevention of patients with ESUS in terms of providing appropriate oral anticoagulation to prevent recurrences of stroke,” he wrote.
“Furthermore, this approach could lead to a paradigm shift in recording sinus rhythm rather than atrial fibrillation on an ECG, with a specific focus on identifying structural changes.”