Anomaly Detection using Multi-channel FLAC for Supporting Diagnosis of ECG

In this paper, we propose an approach for abnormality detection in multi-channel ECG signals. This system serves as front end to detect the irregular sections in ECG signals, where symptoms may be observed. Thereby, the doctor can focus on only the detected suspected symptom sections, ignoring the disease-free parts. Hence the workload of the inspection by the doctors is significantly reduced and the diagnosis efficiency can be sharply improved. For extracting the predominant characteristics of multi-channel ECG signals, we propose multi-channel Fourier local auto-correlations (m-FLAC) features on multi-channel complex spectrograms. The method characterizes the amplitude and phase information as well as temporal dynamics of the multi-channel ECG signal. At the anomaly detection stage, we employ complex subspace method for statistically modeling the normal (healthy) ECG patterns as in one-class learning. Then, we investigate the input ECG signals by measuring its deviation distance to the trained subspace. The ECG sections with disordered spectral distributions can be effectively discerned based on such distance metric. To validate the proposed approach, we conducted experiments on ECG dataset. The experimental results demonstrated the effectiveness of the proposed approach including promising performance and high efficiency, compared to conventional methods.