Predicting Spontaneous Termination of Atrial Fibrillation Based on Analysis of Standard Electrocardiograms: A Systematic Review

Abstract

Background: Forward prediction of atrial fibrillation (AF) termination is a challenging technical problem of increasing significance due to rising AF presentations to emergency departments worldwide. The ability to non-invasively predict which AF episodes will terminate has important implications in terms of clinical decision-making surrounding treatment and admission, with subsequent impacts on hospital capacity and the economic cost of AF hospitalizations.

Methods and results: MEDLINE, EMCare, CINAHL, CENTRAL, and SCOPUS were searched on 29 July 2023 for articles where an attempt to predict AF termination was made using standard surface ECG recordings. The final review included 35 articles. Signal processing techniques fit into three broad categories including machine learning (n = 14), entropy analysis (n = 12), and time-frequency/frequency analysis (n = 9). Retrospectively processed ECG data was used in all studies with no prospective validation studies. Most studies (n = 33) utilized the same ECG database, which included recordings that either terminated within 1 min or continued for over 1 h. There was no significant difference in accuracy between groups (H(2) = 0.058, p-value = 0.971). Only one study assessed recordings earlier than several minutes preceding termination, achieving 92% accuracy using the central 10 s of paroxysmal episodes lasting up to 174.

Conclusions: No studies attempted to forward predict AF termination in real-time, representing an opportunity for novel prospective validation studies. Multiple signal processing techniques have proven accurate in predicting AF termination utilizing ECG recordings sourced from a database retrospectively.

Keywords: atrial fibrillation; electrocardiogram; entropy; frequency analysis; machine learning; prediction; termination; time–frequency analysis.

FIGURE 1

Search flow diagram for studies included in this review. Databases included MEDLINE, EMCare, CINAHL, CENTRAL, and SCOPUS.

FIGURE 2

Accuracy among categories. Box and whisker plot indicating the median (90% for all groups) and interquartile ranges.

FIGURE 3

Machine learning process. Signal preprocessing was variably applied to raw ECG signals to reduce baseline wander, reduce high‐ and low‐frequency noise and/or cancel QRST complexes to isolate the atrial activity (AA). Some groups did not require any signal preprocessing. Signal representation is the process of presenting ECG data or isolated AA into an alternative domain, for example transitioning from the time domain into the frequency domain. This allows extraction of variables which cannot be ascertained from the raw data. Feature extraction is the process of extracting numeric variables from the processed signals for subsequent analysis. Feature selection was either automatic whereby a statistical process was utilized to optimize selection of features to be fed into the machine learning classifier or manual, whereby the authors selected the features themselves. Classification is the stage where previously selected features were fed into a machine learning algorithm which categorized the sample as terminating or non‐terminating.

FIGURE 4

Entropy analysis process. All entropy studies underwent signal preprocessing to produce isolated atrial activity (AA) which could be subjected to entropy analysis. Signal representation included either creating a time–frequency distribution via the use of a transform function or selective filtering of the AA around the dominant frequency to create a graphical representation of the so‐called main atrial wave (MAW). Entropy analysis was then done whereby the approximate, sample or wavelet entropy of the signal was calculated. Classification was then achieved using a threshold whereby signals with entropy above a specific point were classified as non‐terminating and signals below were classified as terminating or using a linear classifier.

FIGURE 5

Time–frequency/frequency analysis process. All time–frequency/frequency analysis studies underwent signal preprocessing to produce isolated atrial activity (AA). Signal representation included either creating a time–frequency distribution or power spectral representation via the use of a transform functions. Features were then extracted from these signals including dominant frequency, major atrial fibrillation frequency, atrial fibrillatory rate and wavelet coefficients vectors. All signals were then classified using a threshold whereby signals with entropy above a specific point were classified as non‐terminating and signals below were classified as terminating however the wavelet coefficients vectors were classified using a central tendency measure.