• Leveraging "Small Data" Through Predictive Analytics To Support Clinical Decision-making

      McKinley, DeAngelo
      ABSTRACT DeAngelo Jamond McKinley, BS LEVERAGING “SMALL DATA�? THROUGH PREDICTIVE ANALYTICS TO SUPPORT CLINICAL DECISION-MAKING Under the direction of Ayman Akil, Ph.D. The United States pays twice as much for healthcare on average than any other nation on Earth, yet the quality of care is not guaranteed. Differences in clinical decision-making (CDM) have been identified as a leading contributor to the variability in the quality of healthcare throughout the United States. The healthcare industry has sought to employ predictive analytics to improve the quality of care. The major focus of these efforts has been centered on supporting resource management and streamlining operational tasks, however, there has been little effort to address the needs of clinicians in direct patient care. A better understanding of the mechanisms behind existing interventions and patient specific factors that lead to poor clinical outcomes are critical to improving quality in healthcare. Our central project aim is to demonstrate how and where small data predictive analytics can be used to support CDM. In specific aim 1 we used univariate regression to evaluate the impact of a pharmacist-led intervention in African American patients with heart failure (HF). We found that the pharmacist-led intervention had a greater impact on HF-related readmission than non-HF related readmission. In specific aim 2 we used a linear model tree to identify patient specific factors predictive of 30-day readmission. We found numerous factors predictive of HF-related and non-HF related readmission with no clear distinction between the predictors of these two clinical outcomes. In specific aim 3 we assessed the relationships between medication adherence, lifestyle modifications, and long-term blood pressure control through univariate regression, and the Andersen-Gill model, a timeseries model for predicting the probability of recurrent uncontrolled hypertensive events. Results suggested that a combination of lifestyle modifications may have a comparable effect on long-term blood pressure control to medication adherence along. In this work we supported CDM by evaluating the effect of interventions on a population of intertest, identifying factors contributing to the risk of an untoward medical event, and codifying significant contributors to a clinical phenomenon. In future work we aim to deploy our HF risk prediction model in a clinical setting and assess the effectiveness of model directed interventions on 30-day HF-related readmission.