An overriding goal of medical care should be tailoring care to individual patient risks, benefits and preferences. To accomplish this, it is first necessary to accurately and reliably estimate patient risk of morbidity and mortality in the absence of treatment. For cardio- and cerebrovascular (CCV) disease, the leading cause of morbidity and mortality in the US, several CCV risk prediction tools exist, however, these tools have substantial shortcomings, including manual entry of risk information, developed on and calibrated to patient populations different than the VA, and do not utilize the full spectrum of clinical data in VA's electronic medical record (VA EMR).
This 1-year pilot project used 5-years of data (2003 - 2007) from 6 VA facilities. The objectives were to: (1) Construct a longitudinal EMR-derived dataset to explore the feasibility of developing an automated risk prediction tool; and, (2) Conduct preliminary analyses to develop a CCV mortality risk prediction model. This project was the first step in a process directed at developing automated tools that can be integrated into the VA EMR or a web-based interface, helping clinicians to optimize and personalize treatment decisions in outpatient settings.
We extracted laboratory and pharmacy data from the DSS National Data Extracts; outpatient visits, inpatient use, ICD-9 codes, and CPT codes from the SAS Medical Datasets; clinical measures from the Corporate Data Warehouse (CDW); and cause of death from the National Death Index (NDI). Our sample included 109,007 male patients who met VA's EPRP "established patient" criteria and had no CCV events in the baseline year; 4,927 had NDI-verified Fatal Cardiac Events during follow-up. Analyses included all patients with a CCV cause of death and 5,000 randomly selected controls. We calculated Framingham risk for CCV death and developed models using Framingham risk factors (blood pressure, age, diabetes, cholesterol) and additional risk factors (hypertension,lipid medications; COPD, sleep apnea, thyroid disease, periodontal disease, inflammatory arthritides). We used 20-fold cross-validation techniques to avoid overfitting.
This project demonstrated the feasibility of obtaining adequate information on patient risk factors. Our results indicated that the Framingham Score performed quite poorly, though recalibrating it to fit the VA population substantially improved the discrimination. Adding medications improved discrimination substantially more and adding other baseline risk factors resulted in further improvement. We also developed a machine-learning model that had slightly better discrimination than the other models, but was less well-calibrated.
The use of multivariable risk/benefit prediction tools to tailor treatments based on patient attributes can increase treatment precision, however, tools to facilitate such improvements in VA are lacking. This project demonstrated the feasibility of developing tools to tailor treatments to individual Veteran circumstances. Future work partnering with Patient Care Services and the Office of Informatics & Analytics will include building and implementing decision support tools to aid communicating and sharing individualized treatments with Veterans, which can help improve outcomes as well as make care more Veteran-centered.
- Sussman JB, Vijan S, Choi H, Hayward RA. Individual and population benefits of daily aspirin therapy: a proposal for personalizing national guidelines. Circulation. Cardiovascular quality and outcomes. 2011 May 1; 4(3):268-75.