A long short-term memory model was able to decipher 12 common heart rhythms using data from 12-lead electrocardiograms (ECGs), according to a study published in the Canadian Journal of Cardiology.
Output of 65,932 digital 12-lead ECG signals from 38,899 patients were collected between 2009 and 2018. The ECG signals were interpreted by cardiologists and stored in the digital core ECG laboratory of China Medical University Hospital. An unbalanced mixture of the 12 heart rhythms were used to train (n=65,932) and test (n=116) the algorithm. Due to the sequential nature of ECG outputs, a 4-layer bidirectional long short-term memory model with 128 neurons was used.
For predicting the 12 heart rhythms, the area for the receiving operator curves (AUC) ranged from a minimum of 0.98 (accuracy, 0.982; precision, 1.000; recall, 0.818) for atrial premature beat to a high of 1.000 for ventricular bigeminy (accuracy, 0.991; precision, 0.833; recall, 1.000), complete heart block (accuracy, 0.991; precision, 0.880; recall, 1.000), sinus tachycardia (accuracy, 0.991; precision, 1.000; recall, 0.900), paroxysmal supraventricular tachycardia (accuracy, 0.991; precision, 0.909; recall, 1.000), and ventricular premature beat (accuracy, 1.000; precision, 1.000; recall, 1.000).
The model had an overall accuracy of 0.90, which was superior to accuracy by internists (0.55±0.14), emergency physicians (0.73±0.08), or cardiologists (0.83±0.10). The model took 5.7 seconds to detect and classify patterns. The clinicians spent more than an average of 30 minutes interpreting the ECG readouts.
The 3 rhythms with the lowest accuracy rate with the computer model (<80%) were also difficult for clinicians to detect. Specifically, ectopic atrial rhythm had the lowest accuracy rates (model, 0.62; internists, 0.0±0.0; emergency physicians, 0.23±0.23; cardiologists, 0.66±0.30), followed by first-degree AV block (model, 0.70; internists, 0.27±0.24; emergency physicians, 0.48±0.20; cardiologists, 0.50±0.15), and atrial premature beat (model, 0.82; internists, 0.37±0.25; emergency physicians, 0.51±0.20; cardiologists, 0.73±0.25).
A major limitation of this study was that ECG noise (eg, baseline wander, electrode motion artifacts, and muscle artifacts) was removed from the input data. It remains unclear how accurate the model would be in a more realistic clinical setting. Furthermore, the accuracy rates were calculated on test set, which for some patterns (complete heart block, ectopic atrial rhythm, and normal sinus rhythm) were less than 10 cases.
“The findings may have clinical relevance for the early diagnosis of cardiac rhythm disorders,” noted the study authors.
Chang K C, Hsieh P H, Wu M Y, et al. Usefulness of machine learning-based detection and classification of cardiac arrhythmias with 12-lead electrocardiograms. Can J Cardiol. 2020;S0828-282X(20)30216-6. doi:10.1016/j.cjca.2020.02.096