Date of Award

5-2024

Document Type

Thesis

Degree Name

Master of Science (MS)

College/School

School of Computing

Department/Program

Computer Science

Thesis Sponsor/Dissertation Chair/Project Chair

Aparna Varde

Committee Member

Hao Liu

Committee Member

George Antoniou

Committee Member

Jiayin Wang

Abstract

This thesis explores a novel approach for early tuberculosis (TB) detection leveraging the rich information within medical big data. We move beyond the traditional reliance on self-reported cough duration and propose a holistic strategy that uses audio data from cough recordings. This comprehensive analysis reveals the influence of factors beyond cough, such as heart rate, in differentiating TB cases. We investigate two primary audio processing methods for TB classification using deep learning models: Mel-spectrograms and Mel-Frequency Cepstral Coefficients (MFCCs). Mel-spectrogram analysis, followed by deep learning models, achieved promising results, but a simpler approach yielded superior performance. Notably, a simple 1D convolutional neural network (CNN) trained on MFCC features achieved an impressive accuracy of 91%, surpassing the World Health Organization's (WHO) requirements for TB screening tests. This finding highlights the potential of MFCC features and 1D CNNs for accurate TB detection using cough sounds. The analysis utilized a large dataset encompassing clinical data from over 1,105 participants with suspected and confirmed TB, along with over 502,252 cough recordings. These findings pave the way for future research aimed at validating this approach on diverse populations and translating it into accessible TB screening solutions, particularly for resource-limited settings. This research underscores the potential of leveraging comprehensive medical big data analysis with audio features for advancements in TB diagnostics.

File Format

PDF

Download

Available for download on Sunday, November 08, 2026

Share

COinS