Graduation Year


Document Type




Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Computer Science and Engineering

Major Professor

Sriram Chellappan, Ph.D.

Committee Member

Ponrathi R. Athilingam, Ph.D.

Committee Member

Kenneth Christensen, Ph.D.

Committee Member

Tansel Yucelen, Ph.D.

Committee Member

Alfredo Weitzenfeld, Ph.D.


Aging, Audio, Healthcare, Heart Disease, Lung Disease


Chronic Obstructive Pulmonary Disease (COPD) and Congestive Heart Failure (CHF) are progressive disorders, and major health concerns among today’s aging population. COPD causes a large mucus buildup in the lungs, leading to chronic cough and difficulty to breathe. CHF causes fluid buildup in the lower lungs due to the failing heart, causing cough and difficulty to breath. People who are clinically diagnosed with COPD or CHF are expected to regularly monitor their symptoms and follow complex medical recommendations in an effort to prevent exacerbation. In this dissertation, we elaborate upon three different machine learning based techniques that we developed for early signs of exacerbation of COPD or CHF symptoms by detecting worsening cough and wheezing.

First, we present the feasibility of leveraging chronic cough samples, recorded using a smart-phone’s microphone, and processing the audio samples via machine learning algorithms, to differentiate COPD from normal (non-COPD) cough patterns. This is done using a cohort of 39 adult cough samples (23 with COPD; 16 healthy patients without COPD which we called "Controls"), evenly spread across both genders, and Random Forests classification algorithm. Next, we propose \emph{TussisWatch}, a smart-phone system to identify cough episodes as early symptoms of COPD or CHF. TussisWatch consists of a two-level, Random Forests classification scheme. At level-one, cough episodes are distinguished as DISEASE (COPD or CHF) or NO DISEASE. If the former is identified, the second-level classifier indicates if the disease is COPD or CHF. If the latter is identified, classification is complete since the user does not have COPD or CHF cough symptoms. TussisWatch was developed with 36 adults cough samples (9 with COPD; 9 with CHF; 18 Controls). Lastly, we consider proper inhaler use among COPD patients, to evaluate the effectiveness of the inhaler in relation to the severity of their symptoms.

In this technique, we test using both cough and breath sound samples collected from 55 clinically diagnosed COPD patients who were hospitalized and were receiving inhaler to manage symptoms. Data was collected before and after proper inhaler administration to determine change using a Support Vector Machine classifier. In all techniques, we extracted commonly used audio features (i.e. Mel-frequency Cepstral Coefficients, Zero Crossing Rate, etc.), and achieved good system performances based on several metrics: Precision, Recall, F1-Scores, Specificity, and Sensitivity. We believe that our proposed systems have the potential to aid early access to healthcare, educate patients on clinically proven self care practices that they can perform at-home and reduce the rates of re-hospitalization caused by COPD exacerbation. In fact, at the end of this dissertation, we present details on the development and future plans on our fourthcoming mobile application to assist COPD patients with at-home symptom monitoring.