Quantitative Computed Tomography Imaging and Machine Learning for Evaluating Chronic Obstructive Pulmonary Disease

Chronic obstructive pulmonary disease (COPD) is characterized mainly by irreversible airflow limitation due to underlying pathological changes such as emphsysema and airway disease. Despite decades of research, there remain major challenges for managing patients with COPD, including: 1) inability to predict episodes of worsening symptoms, known as exacerbations, that may result in emergency room visits or hospitalizations, 2) differentiating COPD from asthma in patients that have features of both diseases, and, 3) prediction of COPD patients at an increased risk of disease progression. Spirometry test measurements are a simple and inexpensive approach to diagnosis COPD and categorize the severity of lung disease. Force expiratory volume in one second (FEV )1is one the most important spirometry test measurement . However, it is a global measurement that does not provide any information about the underlying disease – information that could guide therapy decisions. Computed tomography (CT) imaging, in contrast, provides information about the underlying disease pathology (emphysema, airway disease, etc) and heterogeneity within the lung. In more recent years, predictive models play a significant role in predicting COPD outcomes, such as disease progression, hospitalization, acute exacerbation, emergency room visit and mortality. Machine learning algorithms are widely utilized as models to predict outcomes with maximum accuracy and minimal error. Therefore, the overaching objective of this thesis was to construct a comprehensive feature-set that carries global and regional lung information using combinations of demographics, spirometry test measurements and CT lung features. Then, the machine learning algorithms, including support vector machine (SVM) and neural networks, were applied to predict COPD outcomes, which can be both classification problems such as hospitalization prediction or classification of COPD/asthma, and regression problems such as predicting COPD progression as measured byFEV . Fea1ure selection also plays an important role in identifing the most important features, and for dimensionality reduction to reduce the complexity of the learning algorithm and the probability of overfitting. In this regard, this thesis proposed novel hybrid features selection with the aim of finding the most important predictors. Additionally, a feature selection algorithm based on nonnegative matrix factorization (NMF) with geometry structure preserving and sparsity consideration was proposed to find most important features. In each of our studies, we demonstrated the performance of the learning algorithms were considerably increased by using CT pulmonary imaging features. Key words: COPD hospitalization, Spirometry test, CT lung biomarkers, Machine learning and Feature selection.