Algorithm for Atrial Fibrillation Detection during Biologix Sleep Study

João Pedro Walsh Crema; Diego Munduruca Domingues; Paloma Rodrigues Rocha; Sara Quaglia Campos Giampá; Geraldo Lorenzi-Filho

doi:10.1055/s-0045-1811820

Sleep Science, Table of Contents

CC BY-NC-ND 4.0 · Sleep Sci 2024; 17(S 01): S1-S89
DOI: 10.1055/s-0045-1811820

Sleep Science Supplement 2024

Algorithm for Atrial Fibrillation Detection during Biologix Sleep Study

Authors

João Pedro Walsh Crema

¹Biologix Sistemas S.A., São Paulo, SP, Brazil
Diego Munduruca Domingues

¹Biologix Sistemas S.A., São Paulo, SP, Brazil
Paloma Rodrigues Rocha

¹Biologix Sistemas S.A., São Paulo, SP, Brazil
Sara Quaglia Campos Giampá

²Laboratório do Sono, LIM 63, Divisão de Pneumologia, Instituto do Coração, InCor, Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo, São Paulo, SP, Brazil
Geraldo Lorenzi-Filho

²Laboratório do Sono, LIM 63, Divisão de Pneumologia, Instituto do Coração, InCor, Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo, São Paulo, SP, Brazil

Abstract

Full Text

Keywords

atrial fibrillation - model - neural network - sleep apnea - obstructive

Introduction: Atrial fibrillation (AF) is common in patients with obstructive sleep apnea (OSA). However, AF is not systematically evaluated during sleep studies.

Objective: To develop an artificial neural network (ANN) algorithm for detection of AF in patients with suspected OSA undergoing Biologix sleep study.

Methods: The study was conducted in two parts. In the first part, we used 4 public databases and RR intervals were detected by electrocardiogram (ECG). The k-fold cross-validation method (k=10) was used to develop and validate an ANN algorithm to detect AF from RR intervals. The ANN model produced cardiac rhythm (AF or normal) predictions for every group of 60 heart beats throughout each exam. In the second part of the study, the Biologix oximeter photoplethysmography (PPG) signal was used to detect heart beats. We used data from patients that underwent simultaneous type 1 polysomnography with ECG and Biologix.

Results: In the first part, the model was trained on 18M RR intervals, with 50% being positive for AF. The model was then tested on 36M intervals, from 249 patients, of which 169 (68%) and 9M beats (22%) were positive for AF. The ANN model, evaluated across all folds, had an area under the curve per beat of 0.994 [95% CI: 0.992-0.996]. The sensitivity to detect AF was 92% [95% CI: 86-95%], specificity of 98% [95% CI: 90-99%] and accuracy of 93% [95% CI: 90-96%]. In the second part, 9 patients (5%) had AF confirmed by polysomnography ECG. The mean F1-score for the quality of R peak detection on PPG was 0.95 [95% CI: 0.85-1]. Patient classification using PPG data had sensitivity of 100% [95% CI: 63-100%] and specificity of 100% [95% CI: 97-100%], with McNemar's test yielding p<0.01.

Conclusion: The ANN algorithm developed is accurate in detecting AF using both the ECG RR intervals and the PPG from the Biologix oximetry system. Therefore, the new algorithm can detect AF in patients with suspected OSA undergoing Biologix sleep study.