Sparse Coding for Efficient Biomedical Signal Processing

Project Summary

Research team:

Abstract

The proposed project aims to develop, implement and validate new algorithms for biomedical signal processing and classification, focusing especially on ECG signals, using methods emerging from the theory of sparse coding.

Sparse coding is a promising tool in signal processing, and consists of representing data using a small set of signals from a fixed or adaptive overcomplete dictionary. These compact representations can be exploited in various applications for more efficient acquisition, processing, or serve as features for signal classification. The literature on sparse coding and its applications, especially compressed sensing, is extensive, and there have been many algorithm developed for common tasks like coding and dictionary learning

For ECG signal processing, sparse coding is a promising approach for some emerging application with strict efficiency and accuracy requirements, e.g. wearable medical devices for long-term ECG monitoring, for monitoring driver state for increased automotive safety, and for person identification based on ECG biometric traits. Sparse coding theory is well positioned as a solution for these technical challenges, due to its inherent efficiency in representation and the possibility of learning dictionaries of characteristic features.

The main objectives of the project are to design, implement and evaluate solutions for identifying different pathologies of the heart and for person identification, based on sparse coding of ECG data with dictionaries of characteristic features. The algorithms will be integrated in an open-source software library. The dissemination of the results will be done via publishing in indexed journals and high-quality conferences. The project is planned for 24 months.

Objectives

The main topic of the project is using sparse coding techniques for enhanced processing and classification of medical signals, focusing especially on ECG signals. This involves working along two directions: (i) design good dictionaries for sparse and accurate representation of the data, and (ii) develop algorithms which rely on the sparse representations for improved signal processing tasks, e.g. segmentation, denoising or classification of the data. To this end, we define the following particular objectives:

• O1. Perform review of the state-of-the-art sparse representation models and applications used in medical signal processing, and especially for processing ECG signals
• O2. Design, implement and evaluate a solution for identifying different pathologies of the heart, based on ECG sparse coding using dictionaries of discriminative features
• O3. Design, implement and evaluate a solution for person identification using ECG sparse coding with dictionaries of biometrical traits extracted from the ECG signal acquired from chest, finger or hand palm
• O4. Investigate extensions of the current sparse models: use alternative sparse models, joint analysis of multi-channel ECG signals, or fusing signals acquired from multiple sensors
• O5. Integrate the developed algorithms into an open-source software library and publish it online
• O6. Disseminate the research results through papers at international conferences and JCR indexed journals

Estimated results

• 1 journal paper in a JCR indexed journal.
• 2 conference papers in international conferences.
• Technical reports following the completion of each project stage, and a final technical report at project finalization

Research team

1. Project director: dr. Nicolae Cleju
2. Research mentor: dr. Iulian Ciocoiu