Search

In this study, we propose a post-hoc explainability framework for deep learning models applied to quasi-periodic biomedical time-series classification. As a case study, we focus on the problem of atrial fibrillation (AF) detection from electrocardiography signals, which has strong clinical relevance. Starting from a state-of-the-art pretrained model, we tackle the problem from two different perspectives: global and local explanation. With global explanation, we analyze the model behavior by looking at entire classes of data, showing which regions of the input repetitive patterns have the most influence for a specific outcome of the model. Our explanation results align with the expectations of clinical experts, showing that features crucial for AF detection contribute heavily to the final decision. These features include R-R interval regularity, absence of the P-wave or presence of electrical activity in the isoelectric period. On the other hand, with local explanation, we analyze specific input signals and model outcomes. We present a comprehensive analysis of the network facing different conditions, whether the model has correctly classified the input signal or not. This enables a deeper understanding of the network's behavior, showing the most informative regions that trigger the classification decision and highlighting possible causes of misbehavior.

Aims: Screening for asymptomatic atrial fibrillation (AF) could prevent strokes and save lives, but the AF burden of those detected can impact prognosis. New technologies enable continuous monitoring or intermittent electrocardiogram (ECG) snapshots, however, the relationship between AF detection rates and the burden of AF found with intermittent strategies is unknown. We simulated the likelihood of detecting AF using real-world 2-week continuous ECG recordings and developed a generalizable model for AF detection strategies., Methods and Results: From 1738 asymptomatic screened individuals, ECG data of 69 individuals (mean age 76.3, median burden 1.9%) with new AF found during 14 days continuous monitoring were used to simulate 30 seconds ECG snapshots one to four times daily for 14 days. Based on this simulation, 35-66% of individuals with AF would be detected using intermittent screening. Twice-daily snapshots for 2 weeks missed 48% of those detected by continuous monitoring, but mean burden was 0.68% vs. 4% in those detected (P < 0.001). In a cohort of 6235 patients (mean age 69.2, median burden 4.6%) with paroxysmal AF during clinically indicated monitoring, simulated detection rates were 53-76%. The Markovian model of AF detection using mean episode duration and mean burden simulated actual AF detection with ≤9% error across the range of screening frequencies and durations., Conclusion: Using twice-daily ECG snapshots over 2 weeks would detect only half of individuals discovered to have AF by continuous recordings, but AF burden of those missed was low. A model predicting AF detection, validated using real-world data, could assist development of optimized AF screening programmes., (© The Author(s) 2020. Published by Oxford University Press on behalf of the European Society of Cardiology.)

Aims: Screening for asymptomatic atrial fibrillation (AF) could prevent strokes and save lives, but the AF burden of those detected can impact prognosis. New technologies enable continuous monitoring or intermittent electrocardiogram (ECG) snapshots, however, the relationship between AF detection rates and the burden of AF found with intermittent strategies is unknown. We simulated the likelihood of detecting AF using real-world 2-week continuous ECG recordings and developed a generalizable model for AF detection strategies.Methods and Results: From 1738 asymptomatic screened individuals, ECG data of 69 individuals (mean age 76.3, median burden 1.9%) with new AF found during 14 days continuous monitoring were used to simulate 30 seconds ECG snapshots one to four times daily for 14 days. Based on this simulation, 35-66% of individuals with AF would be detected using intermittent screening. Twice-daily snapshots for 2 weeks missed 48% of those detected by continuous monitoring, but mean burden was 0.68% vs. 4% in those detected (P < 0.001). In a cohort of 6235 patients (mean age 69.2, median burden 4.6%) with paroxysmal AF during clinically indicated monitoring, simulated detection rates were 53-76%. The Markovian model of AF detection using mean episode duration and mean burden simulated actual AF detection with ≤9% error across the range of screening frequencies and durations.Conclusion: Using twice-daily ECG snapshots over 2 weeks would detect only half of individuals discovered to have AF by continuous recordings, but AF burden of those missed was low. A model predicting AF detection, validated using real-world data, could assist development of optimized AF screening programmes. [ABSTRACT FROM AUTHOR]

Introduction: Atrial fibrillation (AF) is the most common serious abnormal heart rhythm, affecting 34M people worldwide. A range of wireless sensors can help detect AF, but the classification of short noisy ECG intervals more common to these types of sensors is an open problem. Hypothesis: Deep Learning (DL) can automatically identify features (representation learning) and correctly detect AF from wireless ECG data, providing a better classification than algorithms with predefined features (expert features). Methods: Single-lead ECG records were acquired using AliveCor devices and are publicly available in physionet.org. Each ECG was labeled by at least three experts into one of four classes: normal sinus rhythm, AF, other arrhythmias, or noisy data. Five different DL architectures were compared: MobileNet, a convolutional neural network for mobile vision applications, was chosen for the comparison with expert features on an independent set not used for training. Results: We analyzed 8,528 ECG records sampled at 300 Hz with durations of 9 to 61 seconds. After an initial pre-processing to remove the baseline wander and normalize the data, features were extracted using two approaches. 1) An expert-based approach, using standard features based on the heart rhythm and morphological characteristics. 2) A representation learning approach, with features automatically extracted by DL. The features extracted were fed into a neural network classifier, with accuracy shown in terms of sensitivity and specificity in the figure. The DL approach can accurately detect AF: for a fixed specificity of 0.95, sensitivity is 0.84 for the expert features and 0.92 for the DL approach. Conclusions: A DL approach is effective in detecting AF in the presence of short noisy signals from mobile sensors, outperforming an expert-based approach. Measurements with both a mobile sensor and clinical device are needed to validate the DL approach for its future use inside and outside the clinic. [ABSTRACT FROM AUTHOR]