Date of Graduation

7-2026

Document Type

Thesis

Degree Name

Master of Science in Electrical Engineering (MSEE)

Degree Level

Graduate

Department

Electrical Engineering and Computer Science

Advisor/Mentor

Saunders, Robert

Committee Member

Wu, Jingxian

Second Committee Member

Jensen, Morten

Keywords

Hypovolemic Assessment, Peripheral Venous Pressure (PVP), total blood volume (TBV), volume-status classification research

Abstract

Accurate assessment of intravascular volume status remains challenging because traditional measurements such as heart rate, blood pressure, physical examination findings, and laboratory values may be subjective, delayed, or insensitive to early volume loss. Peripheral venous pressure (PVP) waveform analysis provides a potential approach for minimally invasive volume-status monitoring because PVP waveforms can be acquired from peripheral venous access and may contain frequency-domain information related to cardiac activity, respiratory modulation, vascular coupling, and intravascular filling. This thesis presents the development and evaluation of a proof-of-concept embedded prototype for PVP waveform acquisition and volume-status classification. The system integrates a Millar MEMS pressure sensor, bridge-balancing circuitry, analog signal conditioning, programmable offset and gain control, STM32-based data acquisition, embedded integral pulse frequency modulation (IPFM) processing, FFT-based feature extraction, elastic-net logistic regression classification, USB communication, and MATLAB-based visualization and data logging. The prototype acquired PVP waveforms at 500 Hz using 8192-sample analysis windows updated every 1000 samples, corresponding to a 16.384-second analysis window and a 2-second classification update period. The embedded processing pipeline was evaluated for real-time feasibility. The IPFM, FFT feature extraction, and logistic regression evaluation workflow required approximately 1.75 seconds per analysis window, which was below the 2-second update period. This demonstrated that the implemented firmware could complete the required signal-processing workflow within the available real-time interval. The classification framework was evaluated using a human hypertrophic pyloric stenosis validation dataset and porcine hemorrhage recordings. In the human validation dataset, the zero-mean heart rate waveform produced the strongest percentage-based performance, with 94.10% balanced accuracy, and the strongest patient-level cross-validation performance, correctly classifying 15 of 18 patients. Porcine hemorrhage analysis demonstrated that the prototype acquired usable PVP-related waveform content during progressive total blood volume (TBV) loss. In primary cross-subject validation, where models were trained on Subject 2 and tested on device-acquired recordings from Subject 3, the strongest MATLAB balanced accuracies were 77.04% at the 10% TBV threshold and 77.13% at the 15% TBV threshold. Percentage-based combined-subject analysis showed stronger within-dataset feature separability, with 13 of 15 best-performing location-threshold configurations exceeding 85% balanced accuracy. MATLAB reference processing and embedded/device reprocessing showed close agreement when matched model coefficients were used, supporting that the embedded implementation reproduced the MATLAB reference workflow. Overall, the results demonstrate the technical feasibility of an embedded PVP acquisition and signal-processing platform for volume-status classification research. However, the porcine dataset was limited, and the classification results should be interpreted as preliminary feasibility evidence rather than clinical diagnostic validation.

Download

Share

COinS