Naslov
Deep Learning Classification Model for Atrial Fibrillation from Multichannel ECG and Validation on Synthetic and Clinical Databases

Autori
Friganović, Krešimir ; Jović, Alan ; Cifrek, Mario

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
2nd Biomedical Engineering and Instrumentation Summit / - Boston (MA), 2021, 7-7

Skup
2nd Biomedical Engineering and Instrumentation Summit

Mjesto i datum
Boston (MA), Sjedinjene Američke Države, 19.04.2021. - 21.04.2021

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
deep learning ; atrial fibrillation ; multichannel ECG

Sažetak
The deep learning approach for improved disorder classification is increasingly used in clinical decision support systems as databases have become more accessible and representative. In this study, we implemented a deep learning model for classifying atrial fibrillation (AF) and normal sinus rhythm (NSR) from multichannel ECG records. Following previous work, we chose the feedforward neural network model consisting of seven dense layers, which has been shown to be effective in classifying AF. A publicly available 12-lead ECG database, the Physionet/Cinc 2020 Challenge, was chosen because it contains clinical data from multiple sources. A balanced dataset of 2603 patients with AF and 3000 subjects with NSR were selected. Additionally, 240 synthetic AF ECG records were used for comparison. The records were preprocessed by filtering, removing DC component, normalizing, and extracting R peaks using the Pan-Tompkins algorithm. RR segments were extracted and resampled to equal size and used as inputs for the deep learning model. The dataset was split patient-wise into training (60%), validation (20%) and test (20%) sets. To take advantage of the multichannel ECG, the voting for majority class was applied to the model. The best F1-score was 95.8% (aVR lead) using the clinical database. For the synthetic database, the best F1-score was 99.8% (II lead). By applying voting to all leads, test results on clinical database were improved by 1.1%. Training the model on the clinical database showed good results (F1-score = 98.3%) when tested on the synthetic database. However, training on the synthetic database gave poor results (F1-score = 64.3%) when tested on clinical database. This study showed that we can achieve better results using multichannel classification. We also showed limitations of using the synthetic database in generalizing ECG changes in morphology during AF episodes.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo

POVEZANOST RADA