AI-ML for quantitative blood pressure and blood glucose estimation from non-occlusive PPG/ECG bio-signals

• Main Author: Vignesh Sujith Menon
• Other Authors: Ng Yin Kwee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Mechanical engineering::Mechatronics
• Online Access: https://hdl.handle.net/10356/168388
• Institution: Nanyang Technological University

Diabetes, a condition caused by high levels of glucose in the blood that is not properly controlled, is a major contributor to human mortality. Even though this disease affects more than 500 million individuals, there is a lack of non-invasive methods for checking glucose levels, making the use of a glucometer with finger pricking the most common and reliable option for daily glucose testing in personal healthcare. Over the last decade, however, there have been notable advances in the development of wearable devices that monitor blood glucose levels continuously. Thus, the prospect of a non-invasive glucose estimation system appears to be on the cusp of becoming a reality. This FYP report serves to ascertain the feasibility of employing Continuous Glucose Monitoring on a prototype wearable device, employing Machine Learning algorithms to facilitate the estimation of Blood Glucose Levels. We have two objectives in this study. The first is to develop a universal windowing algorithm for PPG and ECG signals to be employed on wearables, and the second, is to develop a robust and effective Blood Glucose Estimation algorithm. Electrocardiogram and Photoplethysmogram signals that are collected from participants adorning the wearable device, undergo a series of pre-processing steps before going through a rigorous windowing algorithm to reject noisy and unclean wavelets. Clean wavelets are then passed to the next stage where fiducial points are marked, following which, 9 features are extracted from them and used as inputs to the Machine Learning model. The model then churns out a prediction of the individual’s Blood Glucose Level. The results of this study found that the proposed algorithm can predict an individual’s Blood Glucose Level with a Mean Absolute Error of 4.8mg/dL. Further, the developed windowing algorithm has the potential to be applied to wearable devices to reject noisy and unwanted wavelets based on wavelet shape variation.