Publication Date

Fall 2019

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Nima Karimianbahnemiri

Keywords

deep learning, ecg, electrocardiogram, glucose, hyperglycemia, machine learning

Subject Areas

Computer engineering; Biomedical engineering; Computer science

Abstract

Hyperglycemia is characterized by an elevated level of glucose in the blood. It is normally asymptomatic, except for an extremely high level, and thus a person can live in that state for years before the negative - sometimes irreversible - health impacts appear. Unexpected hyperglycemia can also be an indication of diabetes, a chronic disease that, when not treated, can lead to serious consequences, including limb amputations and even death. Therefore, identifying hyperglycemic state is important. The most common and direct way to measure a person’s glucose level is by directly assessing it from a blood sample by pricking a finger, which causes discomfort and even pain. The constant finger pricking can also lead to bruising and increases the possibility of infection. This thesis presents a non-invasive technique of detecting hyperglycemia by using a person’s electrocardiogram (ECG) and deep learning. The ECG signal is preprocessed to remove noise, identify fiducial points, extract and adjust features, remove outliers and normalize the data. This thesis applied a novel approach to feature extraction in which, instead of just using fiducial amplitudes and intervals, a direct line was drawn between fiducial points and its length and slope were used as features. The labeled features were used in 10-layer deep neural network and resulted in an area under the curve (AUC) of 94.53%, sensitivity of 87.57% and specificity of 85.04%. Such strong performance indicates that ECG carry intrinsic information that can be used to identify hyperglycemic state, enabling the use of ECG-based hardware together with deep learning for non-invasive hyperglycemia detection.

Download

Share

COinS