Energy-Efficient Real-Time Electrocardiography Telemonitoring

To achieve real-time electrocardiography (ECG) telemonitoring, one of the major obstacles to overcome is the scarce bandwidth. Compressed sensing (CS) has emerged as a promising technique to greatly compress the ECG signal with little computation. Furthermore, with edge-classification, the data rate can be reduced by transmitting abnormal ECG signals only. However, the reconstruction algorithm of compressed data is of very high complexity, and thus is infeasible for edge devices to apply classifier on reconstructed data. Compressed analysis (CA) then becomes a challenging problem and requires domain knowledge from both CS and machine learning (ML). There are three main limitations for edge-classification on ECG data: limited amount of labeled ECG data, tight battery constraint of edge devices and low response time requirement. Task-driven dictionary learning (TDDL) appears as an appropriate classifier to render low complexity and high generalization. Intuitively, we first combined CS with TDDL directly (CA-N) and found that classification performance degraded. Consequently, CA-N required higher complexity model, namely, more atoms in the dictionary to bridge the performance gap. In this paper, we proposed an eigenspace-aided compressed analysis (CA-E), which integrated principal component analysis (PCA), CS and TDDL, and sustained not only light complexity but high performance under exiguous labeled ECG dataset.

The contribution of this work comes in two aspects:

• Our proposed CA-E reduces about 67% parameters, 76% training time and 87% inference time and has a smaller performance variance to CA-N counterpart.
• Our proposed CA-E outperforms the conventional support vector machine and dense neural network by about 10% margin when the number of data is halved.

Recommended citation:

K.-C. Hsu, B.-H. Cho, C.-Y. Chou and A.-Y. (ANDY) Wu, ‘‘Low-Complexity Compressed Analysis in Eigenspace with Limited Labeled Data for Real-Time Electrocardiography Telemonitoring,’’ Signal and Information Processing (GlobalSIP), IEEE Global Conference on, Nov. 2018.