Identification of spatiotemporal dispersion electrograms in atrial fibrillation ablation using machine learning: A comparative study. (February 2022)

• Record Type: Journal Article
• Authors: Ghrissi, Amina; Almonfrey, Douglas; Squara, Fabien; Montagnat, Johan; Zarzoso, Vicente
• Is Part Of: Biomedical signal processing and control. Volume 72(2022)Part A
• Publication Date: 2022-02
• Subjects: Atrial fibrillation -- Catheter ablation -- Machine learning -- Multichannel electrogram -- Multipole catheter -- Spatiotemporal dispersion -- Transfer learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103269
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A new patient-tailored ablation protocol using multipole catheters has been proposed. Visual inspection of spatiotemporal dispersion patterns lacks reproducibility. Machine learning models for automatic classification of electrograms are benchmarked. The decision-aid solution can work in real time with moderate computational resources. The solution can improve catheter ablation efficacy, while reducing duration and cost. Abstract: Atrial Fibrillation (AF) is the most widespread sustained arrhythmia in clinical practice. A recent personalized AF therapy consists in ablating areas displaying spatiotemporal dispersion (STD) electrograms (EGM) with the use of catheters. Interventional cardiologists use a multipolar mapping catheter called PentaRay to identify visually atrial sites with STD pattern by visual inspection. In this contribution, we propose to automatize the identification of STD EGMs using machine learning while comparing several features. The aim is to design a data representation and an adapted classification algorithm for accurate STD detection with affordable computational resources and low prediction time. Four data formats are considered: 1) EGM matrices; 2) EGM plots; 3) three-dimensional EGM plots; 4) maximal voltage absolute values. Convolutional neural networks and transfer learning based on the VGG16 architecture are benchmarked. Classification results on the test set show that extracting features automatically with VGG16 is possible and yields comparable results to classifying raw EGM recordings with values of accuracy and AUC of 90%. However, the overall precision and F1 score are low (50%), which can be explained by the high class imbalance ratio. This issue is addressed with data augmentation. Due to its low computational cost, our solution can also be deployed in real time.