Spatial enhancement of ECG using multiple joint dictionary learning. (September 2019)

• Record Type: Journal Article
• Authors: Nallikuzhy, Jiss J.; Dandapat, S.
• Is Part Of: Biomedical signal processing and control. Volume 54(2019)
• Publication Date: 2019-09
• Subjects: 00-01 -- 99-00; Electrocardiography (ECG) -- Derived ECG -- Dictionary learning -- ECG synthesis -- Spatial resolution of ECG; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.101598
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A new personalized dictionary learning model for enhancing spatial resolution of ECG. Joint dictionary learning used to convert low to high-resolution (LR to HR) ECG. Conversion function improves mapping between sparse coefficients of LR and HR ECG. Segmenting and learning ensures diagnostic quality in the estimated ECG. Personalized dictionary implies less over-fitting and good reconstruction quality. Abstract: Transforming a low-resolution electrocardiogram (ECG) recorded using a few electrodes to a high-resolution version can improve the continuous cardiac monitoring environment. Such a transformation can enhance the information available in few leads and increase the spatial resolution. Proper selection of this transformation can provide information related to the heart as good as that obtained while using a large number of leads. This work explores and evaluates a new technique for improving the spatial resolution of a low-resolution ECG by integrating the sparse coding and the joint dictionary learning framework. The information available from the standard twelve-lead ECG subset (low-resolution) is enhanced using a previously learned model for obtaining an estimate of the complete twelve-lead ECG (high-resolution). The model used for transformation is an over-complete dictionary and is learned by using a joint dictionary learning approach. The joint dictionary consists of high and low-resolution dictionaries corresponding to the high and low-resolution ECG and are learned simultaneously. Joint learning of the dictionaries can ensure similar sparse representation for both high and low-resolution ECG. A mapping between the sparse coefficients of high and low-resolution ECG, termed as the conversion function is also learned for improving the reconstruction accuracy. The twelve leads consist of various clinically significant features that differ in amplitude and frequency. In order to preserve these features, the signal is divided into multiple segments and for each segment, a joint dictionary and a conversion function are learned. Standard, as well as diagnostic distortion measures, are used to study the closeness between original and reconstructed ECG lead signals. The diagnostic quality of the ECG obtained using the proposed model is also compared with the existing methods. The analysis of the results shows that the proposed model captures diagnostic content in the spatially enhanced ECG effectively when compared to existing models.