Sensitive and precise detection of ventricular oversensing using deep learning

Background Ventricular oversensing in patients with implantable cardioverter defibrillators and permanent pacemakers can cause inappropriate shocks or inhibition of pacing and non-physiological oversensing is usually the first manifestation of ventricular lead dysfunction. Ventricular oversensing is often incorrectly logged as non-sustained ventricular tachycardia (NSVT) by cardiac devices, but identifying these episodes is time consuming since some episodes are subtle and the overall prevalence of oversensing is low. Purpose To develop a clinically meaningful deep learning model with a high sensitivity for detecting ventricular oversensing amongst NSVT episodes. Methods NSVT remote monitoring transmissions from three French hospitals were used for training and internal validation and data from a fourth hospital served as an external validation cohort. After pretraining on atrial high rate events and synthetic NSVT episodes engineered from periodic electrograms, a deep ensemble classifier was trained to differentiate physiological or non-physiological oversensing from true NSVT. Results The external validation cohort comprised 2856 NSVT episodes (2732 true NSVT (95.6%) and 124 oversensing (4.4%)) with diverse oversensing episodes including non-physiological signals due to lead dysfunction, P wave, T wave and diaphragmatic oversensing (Figure). The deep ensemble classifier correctly identified all 124 oversensing episodes (100%) with only 8 false positive results (0.3%). This yielded sensitivity 100%, specificity 99.7%, positive predictive value 93.9% and F2 score 98.1 for detection of ventricular oversensing. Conclusion A deep learning classifier can reliably differentiate physiological and non-physiological oversensing from true NSVT with very high sensitivity and precision. This can improve early detection of lead dysfunction, prevent inappropriate therapies and reduce the workload of remote monitoring staff.