Automatic & Robust Electrocardiogram Analysis Using Reinforcement Learning

Abstract

The Electrocardiogram (ECG) is a vital diagnostic test that records the signals of the heart. Among the various components of the ECG waveform, the analysis of P waves and PR intervals holds particular importance in assessing cardiac health. The adequate analysis of ECGs requires the ability to recognize both normal P waves and PR intervals as well as common abnormalities that could indicate disease. Accurately interpreting the small sized P wave in ECG signals is challenging due to individual variations and noise interference. Factors such as body position, respiratory cycle, heart rate, and underlying cardiac conditions can influence the shape, duration, and amplitude of P waves, making it challenging to establish a universal standard for interpretation. Most studies in this field have predominantly relied on the MIT-BIH database, which is limited to data from only 48 cardiac patients, potentially hindering the generalizability of their findings. Neural networks like Restricted Boltzmann Machine (RBM) and deep belief networks (DBN) have shown promising recognition accuracy in binary classification of a single lead data. But these are highly sensitive to anomalies and the size of the training set. Even a few artefacts or mislabelled data can significantly impact the classifier's accuracy. This research focuses on applying the Q learning reinforcement algorithm to the four ECG datasets from the PhysioNet/Computing in Cardiology Challenge (CinC). These datasets represent diverse patient groups from China, United States and Germany. The study is a multilabel classification of five distinct ECG beats from 8,867 patients. Each distinct beat morphology corresponds to one of the five cardiac conditions: Normal Sinus Rhythm, Atrial Flutter, Atrial Fibrillation, 1st Degree Atrioventricular Block, and Left Atrial Enlargement. Variations in P waves and PR intervals are analysed on Lead II and Lead V1 of the ECG. By optimising the Q learning agent for 71,672 beat samples with an alpha and gamma rate of 0.001 and 0.9, respectively, the study achieved an average accuracy of 90.4% with a 12% hamming loss. This robust algorithm can classify around 40,000 samples at an average execution time of 0.06 seconds on a CPU with a processing speed of 1.70 GHz and 4 cores. The Q learning agent effectively handles variations related to P waves and PR intervals with higher accuracy and robustness, considering diversity at three levels: demographics, patients, and lead quality. Future studies can explore the integration of this approach with larger datasets, deep learning models, and real-time monitoring systems to enhance the clinical application of P wave and PR interval interpretation.