Extracting deep features from short ECG signals for early atrial fibrillation detection

Abstract

Atrial Fibrillation (AF) at an early stage has a short duration and is sometimes asymptomatic, making it difficult to detect. Although the use of mobile sensing devices has provided the possibility of real-time cardiac detection, it is highly susceptible to the noise signals generated by body movement. Therefore, it is of great importance to study early AF detection for mobile terminals with noise immunity. Extracting effective features is critical to AF detection, but most existing studies used shallow time, frequency or time-frequency energy (TFE) features with weak representation that need to rely on long ECG signals to capture the variation in information and cannot sensitively capture the subtle variation caused by early AF. In addition, most studies only considered the discrimination of AF from normal sinus rhythm (SR) signals, ignoring the interference of noise and other signals. This study proposes three new deep features that can accurately capture the subtle variation in short ECG segments caused by early AF, examines the interference of noise and other signals generated by the mobile terminal and proposes a new feature set for early AF detection. We use six popular classifiers to evaluate the relative effectiveness of the deep features we developed against the features extracted by two conventional time-frequency methods, and the performance of the proposed feature set for detecting early AF. Our study shows that the best results for classifying AF and SR are obtained by Random Forest (RF), with 0.96 F1 score. The best results for classifying four types of signal are obtained by Extreme Gradient Boosting (XGBoost), with overall F1 score 0.88 and the individual F1 score for classifying SR, AF, Other and Noisy with 0.91, 0.90, 0.73, and 0.96, respectively.

Keywords: Data mining; Deep features extraction; Early atrial fibrillation detection; Medical knowledge engineering.