Accurate machine learning classification of atrial high rate events with automated implementation in remote monitoring

Abstract Background Remote monitoring transmissions from pacemakers and implantable cardiac defibrillators for atrial high rate events (AHREs) include many false positive transmissions that are time consuming to manually verify. Identification of true atrial tachycardia/atrial fibrillation (AT/AF) has important implications for patient management. Purpose To develop a machine learning algorithm that accurately classifies true AT/AF, physiological oversensing and non-physiological oversensing (‘noise’). Methods We trained a 2-dimensional convoluted neural network to perform multi-class classification on the processed atrial electrogram (Figure) from a single vendor. Data was divided in an 80:20 ratio (training / test sets). Classified episodes for which the algorithm’s predicted probability exceeded 90% were deemed ‘high confidence’ classifications. Prospective validation was performed with a custom software developed to automatically access new episodes from the remote monitoring interface and implement the algorithm for classification. Results A total of 14029 home remote monitoring alerts for AHREs were included ( 71.2% AT/AF, 21.1% noise, 7.7% physiological oversensing ) . Average accuracy of the algorithm on the test set was 96%, with high performance metrics for both AT/ AF ( Sensitivity 98.8%, Specificity 91.4%, F1 score 97.7% ) and noise (Sensitivity 94.4%, Specificity 98.8%, F1 score 95.0 ). Sensitivity was moderate (74.2%) for oversensing due to misclassification of oversensing as AT / AF. In almost half (47.8%) the cases where oversensing was misclassified as AT / AF, the atrial rate exceeded 150bpm but was technically not fast enough to fill device counters based on the programmed settings. Implementing a high confidence threshold improved the model’s performance (Table) while retaining 91.5% of total episodes. Prospective automated validation via the remote monitoring interface (300 electrograms: 85.6% AT/AF 4.4% oversensing, 10% noise) yielded an overall accuracy of 98.7%. Conclusion Our machine learning algorithm is highly sensitive and precise for detection of true AT/AF and was easily integrated with the existing remote monitoring interface. Implementation of a ‘high confidence threshold’ to triage AHREs to either automated classification or manual review could streamline remote monitoring, reducing workloads and costs and improving patient outcomes. Once generalizability is confirmed with ‘real-time’ validation in a larger cohort, a similar approach could be extended to other vendors.