Prescreening of Tuberculosis Based on Cough Sound Analysis and Clinical Information Using Machine Learning
| dc.contributor.advisor | Dawit Assefa (PhD) | |
| dc.contributor.author | Woineshet Habtom | |
| dc.date.accessioned | 2025-10-22T13:38:23Z | |
| dc.date.available | 2025-10-22T13:38:23Z | |
| dc.date.issued | 2025-10 | |
| dc.description.abstract | Tuberculosis (TB) remains a major global health concern, particularly in low- and middle-income countries (LMICs) where diagnostic resources are limited. TB is an infectious disease caused by bacteria that most often affects the lungs. Early diagnosis of TB is critical for treatment as well as to reduce its transmission. Current TB diagnosis techniques include chest radiography, sputum smear microscopy, tuberculin skin tests, sputum culture, and Gene-Xpert assays, which often require specialized facilities and trained personnel, posing challenges in resource-limited settings. This study develops a prescreening tool for TB detection that leverages cough sound analysis and patient-specific clinical information through a deep learning approach. The dataset used includes 29,768 cough sounds and clinical information from 1,105 individuals collected across seven countries. The developed model, a convolutional neural network (CNN)-based deep learning architecture for binary classification, employs a two-stage design: a base CNN model (CNNbasedNet121) trained on 1D convolutional layers of cough audio records and an enhanced multimodal model (CNN based-Net) that fuses learned audio features with clinical metadata through dense layers to improve performance. Preprocessing was performed, including scaling, categorical encoding, data imputation for clinical data, as well as clipping and padding of audio waveforms, and memory optimization for both clinical metadata and audio recordings. Data visualization techniques such as histograms, box plots, and correlation heatmaps were used to validate data distribution. Evaluation of the model showed promising classification results with AUCROC of 98.59%, 94.22% accuracy, 91.58% recall, 0.8843 Precision, and 0.8998 F1-score. SHAP analysis was applied to interpret model decisions as part of the explainable AI, which revealed that weight, heart rate, smoking history, and night sweat were among the most influential features. In addition, a user-friendly web-based interface was developed using Streamlit to allow real-time TB prediction based on audio uploads and clinical information input. This interface aims to simplify the process for users, making it more automatic, rapid, cost-effective, and accurate to diagnose TB from cough sound and clinical information. To further assess real-world applicability, the deployed model was tested using a separate dataset collected from selected Ethiopian healthcare facilities. Patient cough audio was recorded using a mobile application called Cough Detector and Recorder App, and clinical data was collected through a structured questionnaire. The deployment results confirmed the model’s practical effectiveness in real-world LMIC settings, achieving an accuracy of 83.33% and absensitivity of 77.78%. | |
| dc.identifier.uri | https://etd.aau.edu.et/handle/123456789/7509 | |
| dc.language.iso | en_US | |
| dc.publisher | Addis Ababa University | |
| dc.subject | TB | |
| dc.subject | Cough sound | |
| dc.subject | Clinical information | |
| dc.subject | CNN | |
| dc.title | Prescreening of Tuberculosis Based on Cough Sound Analysis and Clinical Information Using Machine Learning | |
| dc.type | Thesis |