Nelson DB (Minneapolis COE)
We developed and validated prediction rules for mortality and hospitalization for veterans receiving care at the VISN-13, now a part of VISN-23.
This study analyzed a cross-sectional, mailed survey of all veterans with a VISN-13 inpatient or outpatient encounter between 10/1/96 and 3/31/98, where survey responses were validated and supplemented with administrative data. We used cross validation methodology to develop the mortality and hospitalization prediction scales. Specifically, we generated an initial training set comprising 20,000 survey respondents, a second development/refinement set of 10,000 respondents, and a final test set comprising 10,598 respondents. Within the initial training set we implemented a likelihood based model selection process to identify highly predictive measures of five year mortality and hospital admissions considering age, education, ADLs, prior hospitalizations, sex, six self-reported diagnoses from the survey, marital status, race, self-reported smoking status, general health and change in health in the last year as potential explanatory measures. We used predictors identified in this process in a stepwise model selection process with liberal entry and retention criteria to develop an initial candidate model. The model was refined in application to the second development set and then validated in the final test set using measure of AUC and mean square error of prediction verifying that the predictive ability observed in the development sets did not change in application to the validation set.
We identified a scale for predicting 5-year mortality with AUC of approximately 0.8 or greater based on age, general health, smoking, limitations in bathing, eating and walking. The scale for predicting the risk of hospitalization within 5-years had an AUC of approximately 0.74 based on age, general health, smoking, limitation in walking and presence of arthritis, asthma/COPD, diabetes and heart disease. These scales performed as well as scales incorporating SF-12 scores.
These scales performed as well as scales using commonly used measures of health status or quality of life that are not easily administered in clinical settings.
These simple, easily administered prediction rules could be applied in busy clinical settings to identify veterans with high risk of hospitalization and mortality.