Bayesian models for screening and diagnosis of pulmonary disease

Abstract

Pulmonary and respiratory diseases comprise a large proportion of the global disease burden, responsible for both mortality and disability, with the most common ailments being asthma, chronic obstructive pulmonary disorder (COPD), and allergic rhinitis (AR). This burden is especially concentrated in the developing world, where resources for diagnosing these diseases are more limited. In India, COPD recently became the second leading cause of death. Health workers and many general practitioner doctors are not trained to diagnose pulmonary diseases, leading to high rates of misdiagnosis and underdiagnosis. Over the past six years, our group has been developing screening tools for pulmonary disease. We have developed a mobile toolkit that consists of an electronic stethoscope, an augmented reality peak flow meter, and an electronic questionnaire. Previously, logistic regression has been used for modeling pulmonary disease.

However, logistic regression has certain important limitations: it does not model the problem causally, it isn't very flexible, and it doesn't handle missing data well. In this thesis, we propose a Bayesian framework for disease diagnosis in order to mitigate the issues with logistic regression. A Bayesian network model is presented for predicting the probability of specific pulmonary diseases. The network includes three layers consisting of Diseases, Risk Factors, and Symptoms. We then explored two different approaches to constructing the probability estimates and network parameters employed by the model. The first approach derived the network parameters using training data from a clinical study conducted at a pulmonary research hospital (Chest Research Foundation) located in Pune, India. Arriving patients at the clinic were tested using the MIT mobile toolkit and subsequently examined using a complete pulmonary function testing lab, from which a clinical diagnosis was obtained.

Using this data, we built a Bayesian network which was able to accurately detect patients with asthma, COPD, allergic rhinitis, and other pulmonary diseases, with median AUC=0.9 for COPD, AUC=0.92 for Asthma, and AUC=0.89 for Allergic Rhinitis. The Bayesian model was shown to outperform logistic regression in the case of partially missing data. In our second approach, we constructed a Bayesian network with probabilities derived from expert opinions. We surveyed experienced pulmonologists and used their responses to parametrize our model. This model was also able to accurately classify patients with asthma, COPD, allergic rhinitis, and other pulmonary diseases. For future deployment in the field, our Bayesian diagnostic model has been integrated into Pulmonary Screener, a mobile phone application which is used to collect patient data and calculate probabilities of pulmonary disease.

The current work has expanded the previous version of Pulmonary Screener by updating the model structure, improving the workflow, and making the application more intuitive for health professionals. Given the encouraging results of the generalized Bayesian model presented in this thesis, we believe this framework can be a promising approach for creating diagnostic and screening tools for many applications.