Purpose: Clinical prediction instruments generally incorporate clinical data, whereas models derived from administrative data make use of information coded at discharge. We constructed a mortality model derived from highly detailed administrative data acquired during the first 48 hours of admission.
Methods: Our dataset included information on all patients aged ≥18 years with a principal diagnosis of pneumonia or a secondary diagnosis of pneumonia paired with a principal diagnosis of sepsis, respiratory failure/arrest or influenza, who were admitted between 07/01/07 and 06/30/10 to 347 hospitals that participated in Premier's Perspective database. The dataset was divided into a derivation and validation set. We derived an HGLM inpatient mortality model that included patient demographics, co-morbidities, acute and chronic medications, therapies and diagnostic tests administered in the first 48 hours of admission as well as interaction effects. The final model was applied to the validation set.
Results: The dataset included 200,870 patients in the derivation cohort and 50,037 patients in the validation cohort. In the final multivariable model, 3 demographic factors, 27 comorbidities, 40 medications, 8 diagnostic tests and 10 treatments within the first 48 hours were associated with mortality. The strongest predictors of mortality were early vasopressors (OR 1.79), early non-invasive ventilation (OR 1.59), and early bicarbonate treatment (OR 1.70). The model had a c-statistic of 0.85 in both the derivation and validation cohorts. In the validation cohort, deciles of predicted risk ranged from 0.4% to 33.9% with observed risk over the same deciles from 0.1% to 33.4%. 
Conclusions: A multivariable mortality model based on highly detailed administrative data available during the first 48 hours of hospitalization had good discrimination and calibration. The model could be used for risk-adjustment in observational studies.