Applying Causal Inference and Deep Learning to Improve the Accuracy and Equity of Pulmonary Function Test Interpretation.

点击翻译按钮获取中文摘要

Project Summary Race correction plays a central role in pulmonary function test interpretation, resulting in a decrease in the rates at which respiratory impairments are identified in Black patients along with a decrease in the severity of the impairments thus identified. In this way, race correction may promote health disparities by obscuring respiratory impairments which would otherwise have been identified. However, while the effect of race correction on pulmonary function test interpretation is apparent, the downstream clinical consequences of race correction are unknown. No study has assessed the effect of race correction on the clinical management of patients with respiratory disease. At the same time, few resources have been developed to support the race-free interpretation of pulmonary function tests. Deep learning has the potential to meet this need, offering a way to represent pulmonary function on the basis of flow- volume loop morphology without reference to patient race, thus improving both the accuracy and the equity of pulmonary function test interpretation. The aims of this study are to identify the effect of pul- monary function test interpretation on the clinical management of patients with suspected respiratory disease and to use deep learning to develop novel methods for representing pulmonary function in a race-free manner. First, using more than 100,000 pulmonary function tests performed in the Univer- sity of Pennsylvania Health System (UPHS), a regression discontinuity design will be used to estimate the effect of pulmonary function test interpretation on the diagnosis, testing, and treatment of respira- tory disease. This effect estimate will be combined with preliminary work demonstrating the effect race correction has on pulmonary function test interpretation, to arrive at an estimate of the clinical conse- quences of race correction. Second, unsupervised deep learning will be applied to flow-volume loops collected from UPHS and from the National Health and Nutrition Examination Survey (NHANES) and the construct validity of the resultant interpretations will be assessed. This project is supported by the Palliative and Advanced Illness Research (PAIR) Center of the University of Pennsylvania, which has an outstanding track record of advancing the careers of early stage investigators in health services re- search. The candidate will be mentored by a team with expertise in causal inference, machine learning, and health disparities. Experiential training through this project will be supplemented with coursework in causal inference and deep learning. Findings from this work will directly inform the development of an application for a K23 Mentored Career Development Award, which will use mixed-methods to investigate the ways patients and physicians make use of pulmonary function test interpretations, identify actionable clinical phenotypes within pulmonary function test data, and use supervised deep learning to support the identification of these phenotypes when interpreting pulmonary function tests.