(Georgia Institute of Technology, 2015-04-21)
Kumar, Vikas
An important application of predictive data mining in clinical medicine is predicting the disposition of patients being seen in the emergency department (ED); such prediction could lead to increased efficiency of our healthcare system. A number of tools have emerged in recent years that use machine learning methods to predict whether patients will be admitted or discharged; however, such models are often limited in that they rely on specialized knowledge, are not optimal, use predictors that are unavailable early in the patient visit, and require memorization of clinical rules and scoring systems. The goal of this study is to develop an effective and practical clinical tool for identifying asthma patients that will be admitted to the hospital. In contrast to existing tools, the model of this study relies on routine knowledge collected early during the patient visit. While most tools specific to asthma are developed using only a few hundred patients, in this study the records of 9,000+ children seen across two major metropolitan emergency departments for asthma exacerbations are used. An unprecedented amount of 70 variables is assessed for predictive strength and early availability; a novel sequence of methods including lasso regularized logistic regression and a modified "best subset" approach is then used to select the final 4-variable model. A web-application is then developed that calculates an admission probability score based on the patient parameters at the point-of-care. The methods and results of this study will be useful for those aiming to develop similar tools as well as ED providers caring for asthma patients.