Predictive Analytics for Population Surveillance and Personalized Medicine: The Example of Acute Kidney Injury
The boom in electronic health record use, “big data” initiatives, and the evolution of personalized medicine have all combined to create large volumes of patient data with increasing richness. These trends have also contributed to the desire among researchers and clinicians alike to provide tailored patient care recommendations using large patient databases that predict outcomes based on past experiences with similar patients. Predictive analytic tools can approximate some of these relationships for patients, and they may address limitations, such as clinician reminder fatigue, by delivering only highly appropriate and relevant clinical decision support (CDS) reminders.
However, the use of predictive analytics within electronic health records systems for point-of-care CDS and population health have not yet been widely deployed, and their utility is largely untested.1, 2 Predictive analytics pose significant challenges in incorporating both structured and unstructured data and in maintaining the accuracy of models over time and across populations. Furthermore, the viability of predictive analytics depends on building an electronic health record infrastructure that allows patient- and population-level predictive analytics to be embedded and deployed within appropriate user interfaces.
Within the Geriatric Research Education and Clinical Center, our population health informatics research group focuses on the overall “pipeline” of developing applications that utilize natural language processing (NLP) products and structured data variables for use in both traditional and machine learning risk prediction models. These models are leveraged to pursue population and individual health care improvements through data interpretation, visualization, and clinical decision support within automated population surveillance tools, patient panel/cohort dashboards, and reminders used in clinical care.
We are leading two projects to develop risk prediction models aimed at predicting acute kidney injury (AKI). We are conducting both projects in collaboration with the Office of Analytics and Business Intelligence Predictive Analytics Program (Christopher Nielson, Stephan Fihn). AKI is a serious adverse event with links to progressive chronic kidney disease; AKI occurs in 1 to 5 percent of hospitalized patients, 5 to 20 percent of intensive care unit patients, and in 1 to 31 percent of patients following coronary angiography. Inpatient mortality rates range from 15 percent among general ward patients with isolated AKI to 50 percent among ICU patients requiring dialysis.3 Most importantly, a portion of AKI occurs from preventable exposures— thereby offering the opportunity to provide clinical care recommendations aimed at reducing the risk of AKI.
In these projects, we first developed prediction models for hospital-acquired AKI using electronic health record structured data for patients prior to and during the first 48 hours of admission, and are in the process of integrating variables developed from NLP of text notes into the models. The goal of these models is to predict the occurrence of AKI in the seven days following admission. We found that the key clinical challenge is identifying and optimizing patients’ medications, radiology imaging contrast exposures, and hydration status during the critical admission period when most diagnoses occur and the clinical care plan is implemented.
Secondly, we developed prediction models for AKI following coronary angiography using preprocedural clinical information. Approximately 75 percent of patients in VA receive coronary angiography electively and approximately 60 percent of patients undergo this procedure as an outpatient. As a result, a significant opportunity exists to identify patients at high risk for AKI during the pre-procedural evaluation and to optimize their medications, contrast volume, and hydration status prior to angiography.
We are collaborating with the Health Management Platform Systems Facing Team (Michael Rubin) to develop and deploy a general purpose automated surveillance application that can use these models to evaluate risk-adjusted institutional performance and detect centers that have high and low rates of AKI. We plan to conduct detailed chart review in these centers to study the characteristics of workflow and clinical care variation, which will help determine potential targets for point-of-care clinical decision support and best practice recommendations. We are also collaborating with the VA Clinical Assessment Reporting and Tracking (CART) Program (Thomas Maddox, John Rumsfeld) to integrate the models into patient reminders that will recommend preventable risk factor modification among patients at high risk for AKI. The CART has prioritized development of a predictive analytics supported clinical reminder to be embedded within the CART-CL application using the prediction models developed in this work.
Challenges exist in managing data throughput in real-time as well as maintaining the clinical decision support knowledge base, accuracy of the prediction models over time, and acceptable user interface design and workflow integration. In addition, the utility, cost, and safety of these tools as they are integrated into clinical care must be assessed and monitored. The use of predictive analytics in both population health surveillance and in personalized medicine holds promise for improving clinical care.