Fair risk profiles and predictive models for outcomes of obstructive sleep apnea through electronic medical record data

项目总结

PROJECT SUMMARY Obstructive sleep apnea (OSA) is a sleep-related breathing disorder associated with major co-morbidities and is estimated to affect nearly one billion people worldwide. Moreover, there are differences in prevalence, diagnosis rates, and co-morbid outcomes for OSA based on the demographics of a patient, such as age, race, and gender. The diversity of the clinical manifestations, objective measurements, and outcomes – the phenotype – of OSA underscores the opportunity for predictive models to improve care of patients with OSA. Predicting future (i.e. 5- year post-diagnosis) risks of OSA co-morbid outcomes and predicting how different treatments for OSA affect these risks can help clinicians and patients choose the best treatment strategies. Current OSA outcomes research has key limitations. Prior studies have characterized groups of OSA patients that exhibit similar characteristics, referred to as sub-phenotypes of OSA. However, these studies have been limited by analyzing relatively few variables obtainable from questionnaires. To address this limitation, we will use rich longitudinal electronic medical records (EMR) data to characterize OSA sub-phenotypes and to predict OSA outcome risks for individual patients. To extract insights from EMR data, we will leverage modern computational methods based in machine learning (ML). A second major limitation of existing OSA research is worse predictive model performance for some groups. Model biases have real-world negative implications. The ubiquitous STOP-BANG questionnaire used to screen patients for further OSA testing performs worse for women and Asian individuals, leading to potential delayed, under-, or misdiagnosis of OSA in these groups. To address this limitation, this proposed project will assess and mitigate biases present in our predictive models. To better understand patient factors associated with OSA outcomes, this project has two aims. In Aim 1 clustering methods will be applied to identify groups of OSA patients who share similar sub-phenotypes according to combinations of clinical features and objective measurements present in EMR data. Then, sub- phenotypes will be compared by the rates at which they exhibit different OSA outcomes, providing intuition into potential underlying pathophysiologic differences. In Aim 2, ML classifiers will be applied to build and validate algorithmically fair predictive models for future OSA outcome risks as well as effects of OSA treatments. Patient- specific factors that are consistently associated with differences in OSA outcome risks through Aims 1 and 2 will provide both personalized insights into treatment options and stronger evidence of underlying pathophysiology worthy of further investigation.