Jollis and colleagues [1] have identified important inaccuracies in the use of claims data for outcomes research. One might wonder whether including clinical variables in a database would lead to superior results. The Veterans Affairs computer system provides extensive clinical data on all patients. Two of my patients illustrate the difficulties arising from risk adjustment using routine data from a clinical database.

A 62-year-old white man with a primary discharge diagnosis of congestive heart failure was given medications including digoxin, 0.125 mg daily; furosemide, 120 mg twice daily; and benazepril, 20 mg daily. Chest radiographs showed cardiomegaly and pulmonary vascular redistribution. Examinations by multiple-gated acquisition analysis (MUGA) showed a right ventricular ejection fraction of 20%, a left ventricular ejection fraction of 14%, and biventricular hypokinesis.

Another 62-year-old white man with a primary discharge diagnosis of congestive heart failure was given medications including digoxin, 0.125 mg; furosemide, 10 mg twice daily; and benazepril, 10 mg daily. Chest radiographs showed cardiomegaly, a small right pleural effusion, and vascular engorgement. Examination by MUGA showed a right ventricular ejection fraction of 20%, a left ventricular ejection fraction of 16%, and biventricular hypokinesis.

From these two cases abstracted from the Veterans Affairs database, it would be difficult to decide which patient has the worse prognosis. Yet, any clinician seeing the patients could easily decide because of the marked difference in functional capacity. (The first patient had a New York Heart Association functional class of 2, and the second had a functional class of 4.)

The Duke database would readily distinguish these two patients because it was designed for cardiac patients and thus would include this critical clinical variable. A hospital database designed for other purposes rarely provides all the key data items applicable to the outcome of a given condition. It is often those variables that are not routinely available that provide the best information. Routine databases are easily accessible and cheap to analyze, yet useful outcomes research needs to measure the clinically relevant factors, and this requires active data collection specific to the disease of interest.