Performance comparisons between two artificial intelligence based premature ventricular complex detection models using insertable cardiac monitor

Abstract

Abstract Background Premature ventricular complex (PVC) detection using computationally simple techniques has been recently introduced in insertable cardiac monitors (ICMs). An ICM device capable of monitoring PVC burden over a longer period can be a useful tool for evaluating temporal dynamics of PVC burden leading to appropriate and timely clinical interventions in patients with high PVC burden. Objective The study evaluated the performances of two artificial intelligence (AI) based PVC detection models to improve PVC detection accuracy in ICMs. Methods We developed and evaluated two deep learning convolution neural network (CNN) based PVC detection models which were trained and validated on ICM ECG data. The models were developed and validated using ECG strips stored by ICM implanted in real world patients and collected in the deidentified ICM device data warehouse. All ECG strips were manually annotated for PVCs beat-by-beat. For the first model, only the ICM ECG of three consecutive beats was provided as an input feature whereas for the second model, additional features including QRS morphology differences, RR interval differences, QRS amplitude and slope differences between adjacent beats were extracted and provided as input. The trained AI models were tested on a validation dataset and performance metrics including accuracy, sensitivity and specificity metrics were computed and compared between both models. Results The CNN ResNet-18 models with different input features were trained and validated on a dataset of over 90 patient activated ICM ECG episodes which provided a total of 14,354 beats with over 2,200 PVC beats. This dataset provided PVC beats with different morphologies including monomorphic, polymorphic, bigeminy and trigeminy PVCs. The validation dataset consisted of over 4,300 beats including 664 true PVC beats with various morphologies. On the validation dataset, the first model with only ECG as input obtained an accuracy, sensitivity and specificity of 98%, 92% and 99.1% respectively. Whereas for the second model with ECG plus additional features as input, an accuracy, sensitivity and specificity of 98.9%, 97.3% and 99.2% were obtained respectively (Figure 1). Error analysis of both the AI models revealed that the main reason for misidentifying PVC beats was due to the presence of noise around the QRS complex or the T-wave of the beat in the ECG. Conclusion Detection of PVCs utilizing a novel AI algorithm is both feasible and capable of exceeding sensitivity and specificity performance metrics of non-AI based algorithms and AI models with manually extracted ECG features provide better performance for targeted PVC detection in ICMs.