Date of Award
Summer 2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
First Advisor
Ahamed, Sheikh Iqbal
Second Advisor
Madiraju, Praveen
Third Advisor
Khan, Rumi Ahmed
Abstract
At least two billion people are affected by hemoglobin (Hgb), diabetic-related, and other blood-related diseases. Regular clinical assessments of these problems are conducted by analyzing venipuncture-obtained blood samples in laboratories. A non-invasive, cheap, point-of-care, and accurate test is needed everywhere. We started with Hgb measurement, and after an extensive literature survey, we came up with a non-invasive solution with 10-second Smartphone videos of the index fingertips using custom hardware sets to illuminate the fingers. We tested four lighting conditions with wavelengths in the near-infrared spectrum suggested by the absorption properties of two primary components of blood- oxygenated Hgb and plasma. We found a strong linear correlation between our measured and laboratory-measured Hgb levels in 167 patients with a mean absolute percentage error (MAPE) of 5\%. In our initial analysis, critical tasks were performed manually.Now, using the same data, we have automated or modified all the steps. For all subjects, male subjects, and female subjects, we found a MAPE of 6.43\%, 5.34\%, and 4.85\% and mean squared error (MSE) of 0.84, 0.5, and 0.49, respectively. The new analyses, however, have suggested inexplicable inconsistencies in our results, which we attribute to laboratory measurement errors reflected in a non-normative distribution of Hgb levels in our studied patients, as well as excess noise in the specific signals we measured in the videos. To address these problems, we designed a customizable external attachment to the smartphone designed to limit the noise in this system. This attachment is a plastic box with a topside slot to accommodate the smartphone, and internally a 3-pronged- an electrical circuit, a holder box, and a pressure sensor for the fingertip pressing against the camera lens. The attachment is inexpensive, power-efficient, and portable, and the associated software is programmable to acquire optimal PPG signals. Measurement of blood constituents other than Hgb can also use this attachment with slight modifications, which we plan to do for glycated hemoglobin measurement. Towards that end, we have done exploratory work with an unbalanced dataset to detect hyperglycemic states and found an accuracy of 68\% after oversampling with SMOTEENN.