Date of Award
Fall 2020
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Povinelli, Richard
Second Advisor
Ye, Dong Hye
Third Advisor
Corliss, George
Abstract
Benign Paroxysmal Positional Vertigo (BPPV) is a leading cause of dizziness and imbalance that is responsible for one-third of fall incidents. Diagnosis, however, is ridden with uncertainties and errors. This thesis explores various techniques for BPPV predictive diagnosis from a survey study and proposes measures for predictive performance improvement. Patient-facing surveys are established ways of acquiring medical history in clinical settings and, as this thesis demonstrates, are capable of conveying patterns distinguishable for accurate diagnosis. This work begins by discussing BPPV and vestibular disorders in general, and the risks associated with misdiagnosis or elusive diagnosis. Innovative efforts by medical professionals in vestibular therapy for handling the intricacies of diagnosis and clinical protocols are also explained. To predict BPPV successfully, there are distinguishing marks present in a patient’s dizziness episodic history including the frequency and duration of episodes, the specific nature of the dizziness, and the positional trigger. Given these indicators for predicting BPPV, we develop a number of statistical models on a dataset of survey responses acquired from a clinical cohort study. Next, the thesis establishes a connection between the performance limits of the machine learning methods, and the existence of incorrect answers to the survey prompts. By demonstrating that question misinterpretation and ambiguities exist in the cohort study, we show that certain data quality improvement measures have significant influence on classification performance.