Objective. To compare the precision of risk adjustment in the measurement o
f mortality rates using: (i) data in hospitals' electronic discharge abstra
cts, including data elements that distinguish between comorbidities and com
plications; (ii) these data plus laboratory values; and (iii) these data pl
us laboratory values and other clinical data abstracted from medical record
s.
Design. Retrospective cohort study.
Setting. Twenty-two acute care hospitals in St Louis, Missouri, USA.
Study participants. Patients hospitalized in 1995 with acute myocardial inf
arction, congestive heart failure, or pneumonia (n = 5966).
Main outcome measures. Each patient's probability of death calculated using
: administrative data that designated all secondary diagnoses present on ad
mission (administrative models); administrative data and laboratory values
(laboratory models); and administrative data, laboratory values, and abstra
cted clinical information (clinical models). All data were abstracted from
medical records.
Results. Administrative models (average area under receiver operating chara
cteristic curve = 0.834) did not predict death as well as did clinical mode
ls (average area under receiver operating characteristic curve = 0.875). Ad
ding laboratory values to administrative data improved predictions of death
(average area under receiver operating characteristic curve = 0.860). Addi
ng laboratory data to administrative data improved its average correlation
of patient-level predicted values with those of the clinical model from r=0
.86 to r=0.95 and improved the average correlation of hospital-level predic
ted values with those of the clinical model from r=0.94 for the administrat
ive model to r=0.98 for the laboratory model.
Conclusions. In the conditions studied, predictions of inpatient mortality
improved noticeably when laboratory values (sometimes available electronica
lly) were combined with administrative data that included only those second
ary diagnoses present on admission (i.e. comorbidities). Additional clinica
l data contribute little more to predictive power.