Advanced ECG Signal Analysis for Atrial Fibrillation Detection Using an LSTM Network

Atrial fibrillation (AF) is a heart rhythm disturbance affecting approximately 1% of the global population, being the most common cardiac anomaly. Visual inspection of extensive electrocardiograms (ECGs), patient variabilities, and long-term dependencies make manual AF diagnosis difficult and heavily reliant on a cardiologist’s expertise. Aiming to aid medical diagnosis, this study develops: an innovative R-peak detector to correct database annotation inaccuracies, and machine learning classifiers to enhance AF detection from long-term raw ECGs. The R-peak algorithm combines self-convolved windows filtering, signal derivatives, Shannon energy, and the Hilbert transform. Using the MIT-BIH Atrial Fibrillation database, extracted ECG features included RR intervals (RRi) durations, Jitter and Shimmer disturbances, Shannon and logarithmic energy entropies, instantaneous frequencies, spectral entropy and Wavelet Scattering Transform (WST) coefficients. These were evaluated using Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks. To our knowledge, this is one of the first times that acoustically inspired parameters, Jitter and Shimmer, have described ECG signals. The proposed R-peak algorithm achieved a high accuracy of 98.95%. For AF classification, the best accuracy results were 91.96% for MLP and 98.17% for LSTM, which matches typical state-of-the-art values (≈ 98.00%). Findings indicate that RRi remained the most discriminatory feature, while incorporating P- and T-waves entropies and Jitter significantly enhances performance. Besides LSTM’s superior performance, MLP approaches validate that precise feature engineering allows simpler models to approach state-of-the-art results, highlighting their potential on times of deep learning overuse and of searching for explainable AIs in clinical practice.