THE USE OF BAYESIAN PRINCIPLES TO PREDICT THE OPTIMAL REVISIT INTERVAL AND THE CAUSAL RELATIONSHIP BETWEEN REVISIT INTERVALS AND CHRONIC KIDNEY DISEASE FOR MEDICARE PATIENTS WITH TYPE II DIABETES
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
“Evidence-based follow-up intervals have the potential to reduce healthcare costs per person and improve access without compromising or restricting care” In prediction modeling and causality, one of the main goals is to build an algorithm that best represents a dataset. This process first involves the task of selecting features that would best describe the response variable. This paper aims to address the issue of feature selection in causal models by using Bayesian Network principals and the Least Absolute Shrinkage and Selection Operator (LASSO) methodology executed in Standard Query Language (SQL). LASSO is an algorithm-based method and it yet to be executed solely through SQL. In demonstrating the effective use of this method, this study applies it to 2014 to 2016 outpatient healthcare utilization data from the Centers for Medicare and Medicaid Services (CMS) in order to predict the optimal revisit interval (RVI) and to determine the causal relationship between RVI and Chronic Kidney Disease (CKD) for patients with Type II Diabetes. CKD and diabetes are the 9th and 7th leading causes of death in the United States (respectively), and therefore the cohort of interest in this study. In this study, Likelihood Ratios were used to determine feature importance as they relate to CKD. The RVIs were then calculated as the number of days between two consecutive appointments by the same patient and the same provider. From there causality was derived from determining correlations, sequences, mechanism and counterfactuals related to the relevant features. The optimal RVIs were then deduced from the probabilities of CKD occurring given the presence of specific comorbidities. The probabilities were calculated by analyzing the set of comorbidities that patients had prior to a set date, and the prevalence of CKD after the set encounter date. Results showed that there were 136 million outpatient observations for patients with Type II Diabetes. This resulted from approximately 800,000 distinct patients. The average RVI was 39.45 days, with a median of 91 days with a maximum of 363 days and a standard deviation of 64.3 days. Table 9 includes data on optimal RVI based on various comorbidities. If a patient had a probability of developing CKD that is above 0.5, then their optimal RVI was shorter, compared to those patients with probabilities that are below 0.5. Blood toxicity, orthopedic injuries, anemia and other diseases of the connective tissue were the leading causes and predictors of CKD. The biophysical mechanisms between CKD as a result of kidney overuse due to filtering toxins in the blood, drugs and medications is well known; however, patients who present with a history of the comorbidities should potentially be screened early for CKD as the Likelihood of occurrence may be higher in those patients. Optimal RVI can help ensure that patients with these risk factors are seen before their disease progresses. This method is executable solely in SQL and therefore can be used directly in an Electronic Health Record (EHR) as a decision-making tool for providers. Since it does not involve exporting data from an EHR into statistical tools, patient data is protected, and the process is less time consuming. This method can potentially enable providers identify patients who are at higher risk of developing CKD and be able to allot an optimal RVI in patients need to be seen. Ultimately this can help improve health outcomes for diabetic patients and be leveraged for use with other chronic diseases such as hypertension. Key words: Bayesian Networks, Causality, Revisits, Diabetes, Chronic Kidney Disease, Causality, high-dimensional networks, LASSO Regression