Develop and validate a pneumonia mortality model using automated laboratory data (LAB) and supplementing with Uniform Billing data for demographics, discharge status, and comorbidities. The significance of vital signs (VS) and altered mental status (AMS), which were not automated, was also analyzed.
A model was derived from 16,562 (1,035 deaths) pneumonia admissions across 44 hospitals that exported lab data to the Atlas TM system (CHDCI) in 2000-01. Age, LAB and VS on admission, comorbidities (identified using 6th digit ICD-9 coding) were entered into logistic regressions. ROC assessed model fit and Bootstrapping validated the model internally. Manually abstracted data (n = 77,849, 4,865 deaths) from 184 hospitals provided external validation. Results are presented as odds ratios and 95% confidence intervals.
Median age was 76 and crude mortality was 6.3%. Significant predictors (p < .05) included age, albumin < 2.5 g/dl (1.9, 1.5-2.4), pH arterial < 7.2 (3.1, 2.1-4.7), pH arterial 7.2-7.3 (2.3, 1.6-3.3), BUN > 55 mg/dl (2.7, 2.1-3.3), BUN 40-55 mg/dl (2.0, 1.6-2.5), WBC < 4.4 k/mm3 (1.8, 1.4-2.4), troponin I > 1.5 ng/ml or CKMB > 15 ng/ml (2.3, 1.7-3.3), systolic BP < 100 mmHg (1.5, 1.3-1.8), temperature < 96 F° (1.9, 1.4-2.5), respirations > 39 (1.9, 1.6-2.4), AMS (2.5, 1.9-3.1), and metastatic cancer (2.2, 1.6-3.0). The ROCs for the derivative and validation models were .82 and .81 respectively.
Admission LABs indicating acidosis/alkalosis, hypoalbumenia, leucopenia, cardiac ischemia, and renal dysfunction predict pneumonia mortality. VS & AMS are highly significant and should be automated. Pathophysiological variables commonly measured on admission can generate a parsimonious and clinically plausible predictive model.
Predictive mortality models on pneumonia using clinical data have not been widely adopted due to requirement of costly chart abstraction. This model, based mainly on automated data, is cost effective to implement for prospective pneumonia management and comparative outcome reporting.
Y.P. Tabak, None.