Multi-Stage Detection of Atrial Fibrillation in Compressively Sensed Electrocardiogram

Abstract

Atrial Fibrillation (AF) is a cardiac condition that can be asymptomatic and can lead to increase risk of stroke, heart attack, or death. Long term monitoring of ECG is typically used to diagnose AF. However, long term monitoring of ECG generates a large amount of data that can increase power consumption, storage requirements, and wireless transmission bandwidth. Compressive Sensing (CS) is a compression technique that reduces the amount of data collected and the power consumption of ECG recording devices. However, reconstruction of compressively sensed ECG is a computationally expensive technique. This paper proposes a two-stage AF detection system that detects AF in the compressed domain and only reconstructs ECG segments with low detection confidence to confirm the detection of AF. The system was tested using the Long-Term Atrial Fibrillation Database (LTAFDB) available on Physionet. The system is based on Random Forest built using features extracted using discrete cosine transform, statistical methods, empirical mode decomposition, and wavelet transform. The system achieved an area under the curve (AUC) of receiver operator curve of 0.95 at 50% and 75% compression. The weighted average precision (AP) was 0.94 at 50% and 75% compression, and the F1 score was 0.90 and 0.91 at 50% and 75% compression, respectively. The system was tested using 10-fold record-based cross-validation. Confirming AF detection by reconstructing ECG where AF was detected with low confidence has improved AP, AUC, and F1 score over using an AF detector in the compressed domain only while judicially increasing usage of computational resources.