| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Briefs |
1 Department of Epidemiology and Preventive Medicine, Sackler Medical School, Tel Aviv University, Tel Aviv, Israel;2 Internal Medicine B, Laniado Hospital, Netanya, Israel;
aaddress correspondence to this author at: Department of Epidemiology and Preventive Medicine, Sackler Medical School, Tel Aviv University, Ramat Aviv, Israel; e-mail froom{at}maaganm.co.il
Abstract
Background: The aim of this study was to explore whether electronically retrieved laboratory data can predict mortality in internal medicine departments in a regional hospital.
Methods: All 10 308 patients hospitalized in internal medicine departments over a 1-year period were included in the cohort. Nearly all patients had a complete blood count and basic clinical chemistries on admission. We used logistic regression analysis to predict the 573 deaths (5.6%), including all variables that added significantly to the model.
Results: Eight laboratory variables and age significantly and independently contributed to a logistic regression model (area under the ROC curve, 88.7%). The odds ratio for the final model per quartile of risk was 6.44 (95% confidence interval, 5.42–7.64), whereas for age alone, the odds ratio per quartile was 2.01 (95% confidence interval, 1.84–2.19).
Conclusions: A logistic regression model including only age and electronically retrieved laboratory data highly predicted mortality in internal medicine departments in a regional hospital, suggesting that age and routine admission laboratory tests might be used to ensure a fair comparison when using mortality monitoring for hospital quality control.
Monitoring hospital mortality rates is used to evaluate and improve the quality of healthcare (1) and is one measure of the performance monitoring systems required by hospital-accrediting organizations for quality control (2). This process requires adjustment for differences in severity of illness and other risk factors to ensure a fair comparison. Many attempts worldwide have used readily available independent variables to predict in-hospital mortality (3)(4)(5). Among these variables are the diagnostic codes listed in the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), which are often used for billing. These codes can only be applied retrospectively because they are based on the final discharge diagnosis. This method, however, has been criticized because, although widely available and inexpensive, it lacks clinical details necessary to permit adequate adjustment for each patient’s underlying medical condition (6)(7) and does not differentiate between diseases on admission and hospital-acquired complications, such as shock, that often precede death. Risk-adjustment models that include hospital-acquired complications could therefore overestimate the predictive value of the model and could mask inadequate care by increasing the measured risk of patients whose health deteriorated during hospitalization. In fact, Pine et al. (1), using ICD codes after excluding diagnoses that may have been hospitalacquired, found that the mean areas under the ROC curves in a logistic regression model decreased from 0.87 to 0.75.
Electronically retrieved laboratory data might be useful in predicting mortality because the tests are done routinely on admission, are unbiased by clinical evaluation, and reflect, at least to some extent, disease severity. We explored whether electronically retrieved laboratory data not requiring any data abstraction could predict mortality in internal medicine departments in a regional hospital.
All patients presenting to the internal medicine emergency room at Laniado Hospital over a 1-year period (starting from January 3, 2003) were included in the cohort. Of 23 397 patients, 10 308 (44.1%) were hospitalized in 4 internal medicine departments, which included intensive care patients (nonsurgical patients). All results of clinical chemistry and hematology tests done on admission are stored electronically. There is no hospital policy for routine admission testing, but nearly all patients have a complete blood count and basic clinical chemistries (Table 1A
). Complete blood counts were performed on 2 Advia 120 Hematology Systems (Bayer), and clinical chemistries were performed on 3 VITROS instruments (250 and 750; Ortho-Clinical Diagnostics). The performance of these analyzers is harmonized, and we use internal and external quality-control programs. Administrative data included age and sex. During the 1-year period, there were 573 patient deaths (5.6%).
|
The patients were divided into 2 groups: those who died and those who were discharged alive. Eighty percent of patient deaths occurred within 2 weeks of admission and 94% within 30 days. No follow-up was done after discharge. We compared patient age and sex as well as laboratory results on admission to the hospital between the 2 groups.
To validate the results, we also obtained data from the subsequent year. During that year, of 21 320 patients seen in the emergency room, 10 257 patients (48.1%) were hospitalized and there were 488 deaths (4.8%).
Finally, we reviewed 1284 consecutive charts from those hospitalized in 2 representative departments during the first year and extracted the main discharge diagnosis. Patients were divided into 2 groups according to a diagnosis with or without high hospital mortality (
20%). High-mortality diagnoses included septic shock, urosepsis, various cancers, pneumonia, aspiration pneumonia, and the need for cardiovascular resuscitation on admission. Common lower risk diagnoses included nonspecific chest pain or angina pectoris, myocardial infarction, congestive heart failure, chronic obstructive lung disease, and urinary tract infections.
Statistical analyses included the calculation of proportions for discrete variables and standard deviations for continuous variables. For comparisons between groups, we used the Student t-test for parametric variables and the 
2 test for nonparametric variables. For multivariate analyses, we used logistic regression with all variables that significantly (P <0.05) added to the model retained. All variables not adding significantly to the model were removed together and then added back one at a time to determine whether individually they contributed significantly to the model. For the final model, we calculated the area under the ROC curve. To compare the strength of associations of the various variables, we divided their values into quartiles before logistic regression.
The patients who died were older than the discharged group, but there was no difference in sex (Table 1A
). All of the chemistries and blood count results were significantly different and in the expected direction between the 2 groups. These findings were independent of whether the discharge diagnosis had a high risk of mortality (results not shown). A logistic regression model including only those patients with complete data (n = 9150) was excellent in predicting patient mortalities, with an area under the ROC curve of 90.4%, and included age, 7 chemistry variables, and 5 variables in the complete blood count.
To rule out selection bias and include all admitted patients, we gave all missing data the value of 2.5, whereas the quartiles were given values of 1 to 4. Eight laboratory variables and age significantly and independently contributed to a logistic regression model (Table 1B
; area under the ROC curve, 88.7%). The odds ratio for the final model per quartile of risk was 6.44 (95% confidence interval, 5.42–7.64). For age alone, the odds ratio per quartile was 2.01 (95% confidence interval, 1.84–2.19; area under the ROC curve, 69.6%).
During the subsequent year, the model was also excellent in predicting patient mortalities, with the same variables entering the logistic regression model and an area under the ROC curve of 86.8%. The strengths of association were nearly identical between the 2 years (results not shown).
For all variables that added significantly to the logistic regression model, mortality was lowest in the lowest quartile, highest in the highest quartile, and intermediate for the second and third quartiles (P <0.001 in all cases; results not shown).
The major finding of our study is that a logistic regression model including only age and electronically retrieved laboratory data not requiring expensive data abstraction or diagnostic categories highly predicted mortality in internal medicine departments in a regional hospital. This suggests that routine admission laboratory tests can be used to ensure a fair comparison when using mortality monitoring for hospital quality control. In fact, the final model was comparable to reported models that used a combination of restricted diagnoses and laboratory test results (1). We found an area under the ROC curve of 90.4% for all internal medicine patients admitted to the hospital, whereas Pine et al. (1), who included only patients admitted with a diagnosis of acute myocardial infarction, cerebrovascular accident, congestive heart failure, or pneumonia, reported a combined model based on clinical diagnosis and hand-retrieved laboratory results with an area under the ROC curve of 86%.
We are unaware of previous studies using only routine laboratory data and age to predict mortality in hospitalized patients. Our results, however, are consistent with those of Pine et al. (1), who found that laboratory data added significantly to a restricted administrative model whereas the inclusion of clinical data elements extracted from the patients’ records, including chest radiographic and electrocardiographic findings, mental status, and vital signs, added little to the combined laboratory and restricted administrative model (excluding diagnoses that might have resulted from hospital complications). Our findings suggest that results considering only common conditions can be generalized to all patients with high- or low-risk admission diagnoses hospitalized in internal medicine departments, but not necessarily to patients in surgery departments.
Several methodologic limitations of our study should be acknowledged. This study was done in a single hospital, and extrapolation to other hospitals in Israel and in other geographic areas should be done with caution. Our overall hospital mortality was 5.6%, consistent with other studies (8), but rates have been shown to be dependent on admission diagnosis and the need for surgery, with a 5.3% rate in patients with coronary artery disease and 18.6% in those grouped as having nutritional deficiencies. There also might be socioeconomic or ethnic differential biases. Ethnic origins have been shown to influence mortality rates in patients admitted to 30 hospitals in northeast Ohio with 6 selected diagnoses (9). African-American patients had a 13% lower mortality rate than did white patients (9). However, the large differences between the various laboratory test results between those who died and those who survived the hospitalization, the fact that the differences were always in the expected direction, the consistency of results in those with a higher and lower risk for mortality on admission based on admission diagnosis, and the fact that the results were confirmed from data obtained from a subsequent year suggest that our results can be extrapolated to other settings. Mortality rates still might differ in hospitals that send patients home to die or refer sicker patients to other hospitals. In our setting, referrals to other hospitals for patients on the medical service are extremely rare, as is sending patients home to die. Further studies are warranted in other settings to confirm our findings.
The use of laboratory data to predict overall hospital mortality is not meant to aid the physician because treatment is disease specific. However, our findings have important implications for public policy. We have shown that the risk of dying during hospitalization can be predicted by age and electronically retrieved routine laboratory data without the need to acquire clinical information from a patient’s medical records. This suggests that accurate comparisons of risk-adjusted hospital mortality rates can be made relatively inexpensively to monitor hospital performance and, if our results are confirmed, can also be used in hospitals internally studying comparative risk-adjusted mortality rates between departments.
References
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
