Author ORCID Identifier:

https://orcid.org/0009-0001-7402-8212

Date of Graduation

12-2025

Document Type

Thesis

Degree Name

Master of Science in Electrical Engineering (MSEE)

Degree Level

Graduate

Department

Electrical Engineering

Advisor/Mentor

Wu, Jingxian

Committee Member

Jensen, Morten

Second Committee Member

Saunders, Robert

Keywords

dehydration detection; elastic net logistic regression; frequency component analysis; hemorrhage detection; integral pulse frequency modulation; peripheral venous pressure

Abstract

This thesis studies the effectiveness of utilizing a modified integral pulse frequency modulation (IPFM) algorithm to synthesize peripheral venous pressure (PVP) waveforms for the detection of several physiological conditions using logistic regression. PVP waveforms are collected from 18 human patients and 4 porcine subjects. Human data are used to train models for determining the hypovolemic status of the patient, while the porcine data are used to train models for determining level of anesthesia and detecting internal hemorrhaging. In the human dataset, the waveforms collected from the 18 patients are classified into two groups according to serum chloride levels: resuscitated (≥100 mmol/L) and hypovolemic (< 100 mmol/L). All patients received a fluid bolus of 20 cc/kg, and PVP waveforms were collected both before and after administration to study its effects. The pig dataset contains waveforms collected from the 4 pigs at a total of 12 anesthetic levels, corresponding to varying dosages of isoflurane and propofol. Isoflurane and propofol are administered both before and after controlled hemorrhaging is initiated, and the effects of anesthesia and bleeding are studied in isolation during separate tests. The IPFM model is developed to capture the physiological dynamics of the PVP waveforms, such as the modulation of respiration rate and heart rate over the venous pressure. The IPFM model takes raw PVP waveforms as inputs, and the process generates a synthesized PVP waveform and a synthesized heart rate waveform. It was hypothesized that synthesized heart rate waveforms, generated during IPFM synthesis, may contain all the relevant information needed to correctly categorize a patient or subject, as vasodilation is a primary reaction in all tests conducted, which directly effects the circulation of blood flow from the heart. Raw and synthesized waveforms are windowed in the time domain and converted to their frequency-domain representations to be used as training and testing data for logistic regression models. An elastic net logistic regression algorithm is trained by using the frequency domain windows of the three different types of waveforms, namely, raw PVP, synthesized PVP, synthesized heart rate, to evaluate their impacts on the accuracy of dehydration, anesthetic level and bleeding state detection. PVP waveforms are shown to be excellent indicators of various physiological processes when training logistic regression models, even when using unprocessed signals. Results found that the model trained on frequency windows from IPFM-synthesized waveforms produced the best results overall, with either synthesized PVP or synthesized heart rate waveforms producing the highest accuracy models depending on the test conducted. It is shown that even partial processing using the IPFM algorithm, as is the case with generating estimated heart rate waveforms, provides substantial performance gains over using raw PVP waveforms in training a logistic regression model. The experimental results indicate that the IPFM-synthesization process can be used to remove superfluous information irrelevant to dehydration, anesthetic and hemorrhaging detection, thus improving predication accuracy.

Download

Share

COinS