Creation and implementation of artificial intelligence models in remote monitoring to reduce false positive atrial tachycardia alarms of pacemakers and implantable cardiac defibrillators
L Van Krimpen, A John, A Thiyagarajah, B Sacristan, J Duchateau, T Carbonati, R Dubois, S Ploux, P Bordachar, M StrikAbstract
Background
Remote monitoring (RM) allows quick diagnosis and intervention of device- and health-related complications in patients with a pacemaker or implantable cardiac defibrillator (ICD), but comes with a huge data transmission that increases the workload of healthcare professionals. Automatic assessment and feedback of false positive RM alarms can reduce the RM workload and improve patient security.
Purpose
Reduce RM workload by filtering out false positive Atrial Tachycardia or Atrial Fibrillation (AT/AF) alarms of Biotronik pacemakers and ICDs using Artificial Intelligence (AI).
Methods
Four hospitals provided AT/AF episodes and two RM experts independently labelled the episodes into true AT/AF, noise, and oversensing. Conflicts were settled by a third RM expert. Two residual network (Resnet) models were explored to classify AT/AF alarms in true AT/AF, noise, and oversensing categories. Atrial and ventricular intracardiac electrograms (EGM) were used to train the Resnet EGM model. The Resnet EGM+marker model used the same input but also included EGM marker information. The models were trained, optimized, and validated on the episodes of three hospitals (80% train set, 20% validation set) and the episodes of the fourth hospital formed the external test set. The models were trained to maintain a high sensitivity for AT/AF episodes to avoid missing a true AT/AF event. Model performance was evaluated on the test set by comparing the classification of the models to the labels of the RM experts (gold standard). Lastly, the models were clinically tested for two weeks by running the models on a chrome plug-in a hospital, as shown in figure 1. Model performance is measured with the f1-score, which provides a balance between the positive predictive value and sensitivity, to adequately assess model performance on imbalanced data.
Results
The train/validation set consisted of 8892 episodes of 911 patients, the test set consisted of 1858 episodes of 237 patients, and the clinical data of 307 episodes of 68 patients. Distributions of true AT/AF (86%), noise (13%), and oversensing (1%) were similar in both train/validation and test set. The two-week clinical data consisted of 291 AT/AF (94.8%), 12 noise (3.9%), and 4 oversensing (1.3%) episodes. The Resnet EGM model performed very well and achieved F1-scores of 99.1% (AT/AF), 96.0% (noise), and 73.3% (oversensing) on the test set. On the two-week clinical dataset, the Resnet EGM obtained F1-scores of 99.0% (AT/AF), 73.7% (noise), and 88.9% (oversensing). The Resnet EGM+Marker model did not outperform the Resnet EGM model, showing that adding marker information does not lead to increased performance.
Conclusion
Resnet models filter out false positive alerts while maintaining a high sensitivity for episodes of true AT/AF. This shows the promise of AI models in remote monitoring to reduce the workload and encourages further research.