Research Article
Comparative Study of Machine Learning Methods on Spectroscopy Images for Blood Glucose Estimation
@INPROCEEDINGS{10.1007/978-3-031-34586-9_5, author={Tahsin Kazi and Kiran Ponakaladinne and Maria Valero and Liang Zhao and Hossain Shahriar and Katherine H. Ingram}, title={Comparative Study of Machine Learning Methods on Spectroscopy Images for Blood Glucose Estimation}, proceedings={Pervasive Computing Technologies for Healthcare. 16th EAI International Conference, PervasiveHealth 2022, Thessaloniki, Greece, December 12-14, 2022, Proceedings}, proceedings_a={PERVASIVEHEALTH}, year={2023}, month={6}, keywords={Non-invasive monitoring spectroscopy machine learning blood glucose concentration}, doi={10.1007/978-3-031-34586-9_5} }- Tahsin Kazi
Kiran Ponakaladinne
Maria Valero
Liang Zhao
Hossain Shahriar
Katherine H. Ingram
Year: 2023
Comparative Study of Machine Learning Methods on Spectroscopy Images for Blood Glucose Estimation
PERVASIVEHEALTH
Springer
DOI: 10.1007/978-3-031-34586-9_5
Abstract
Diabetes and metabolic diseases are considered a silent epidemic in the United States. Monitoring blood glucose, the lead indicator of these diseases, involves either a cumbersome process of extracting blood several times per day or implanting needles under the skin. However, new technologies have emerged for non-invasive blood glucose monitoring, including light absorption and spectroscopy methods. In this paper, we performed a comparative study of diverse Machine Learning (ML) methods on spectroscopy images to estimate blood glucose concentration. We used a database of fingertip images from 45 human subjects and trained several ML methods based on image tensors, color intensity, and statistical image information. We determined that for spectroscopy images, AdaBoost trained with KNeigbors is the best model to estimate blood glucose with a percentage of 90.78% of results in zone “A” (accurate) and 9.22% in zone “B” (clinically acceptable) according to the Clarke Error Grid metric.
