Predicting in-hospital mortality. A comparison of severity measurement approaches

Hospital mortality statistics derived from administrative data may not adjust adequately for patient risk on admission. Using clinical data collected from the medical record, this study compared the ability of six models to predict in-hospital death, including one model based on administrative data (age, sex, and principal and secondary diagnoses), one on admission MedisGroups score, and one on an approximation of the Acute Physiology Score (APS) from the revised Acute Physiology and Chronic Health Evaluation (APACHE II), as well as three empirically derived models. The database from 24 hospitals included 16,855 cases involving five medical conditions, with an overall in-hospital mortality rate of 15.6%. The administrative data model fit least well (R-squared values ranged from 1.9-5.5% across the five conditions). Admission MedisGroups score and the proxy APS score did better, with R-squared values ranging from 4.9% to 25.9%. Two empirical models based on small subsets of explanatory variables performed best (R-squared values ranged from 18.5-29.9%). The preceding models had the same relative performances after cross-validation using split samples. However, the high R-squared values produced by the full empirical models (using 40 or more explanatory variables) were not preserved when they were cross-validated. Most of the predictive clinical findings were general physiologic measures that were similar across conditions; only a fifth of predictors were condition-specific. Therefore, an efficient approach to risk-adjusting in-hospital mortality figures may involve adding a small subset of condition-specific clinical variables to a core group of acute physiologic variables. The best predictive models employ condition-specific weighting of even the generic clinical findings.