Spatially Enhanced ECG using Patient-Specific Dictionary Learning

A new method for reconstructing the twelve standard ECG lead signals from its subset using a patient-specific dictionary is discussed and evaluated. The spatial resolution of a standard twelve-lead ECG is high compared to its subset and hence can be considered as a high-resolution ECG. For generating the high-resolution ECG from its subset (i.e., the low-resolution ECG), two patient-specific over-complete dictionaries representing the high and low-resolution ECG are jointly learned. Once the joint dictionaries are learned, sparse coefficients of low and high-resolution ECG can be obtained and a mapping between them can be learned. This mapping, defined as the conversion function, ensures good similarity between the sparse coefficients of low and high-resolution ECG thereby improving the reconstruction accuracy. The proposed model uses a four-lead subset of the standard twelve-lead ECG (high-resolution ECG) as the low-resolution input. The joint dictionaries are learned using the K-SVD algorithm and the sparse coefficients are obtained using the orthogonal matching pursuit (OMP) algorithm. Popular diagnostic distortion measures are used to analyze the diagnostic resemblance between original and reconstructed lead signals and the performance of the model is compared with the existing approaches. The analysis of the performance evaluation results shows that the proposed approach is capable of improving the spatial resolution of ECG without compromising the diagnostic information.