OBJECTIVES: Acute renal failure (ARF) is a prevalent and often preventable adverse drug event. Automated methods for ARF identification facilitate quality improvement and outcome research, but traditional reliance on ICD9 codes has been shown to underestimate the incidence. This study aimed to develop and validate automated algorithms for the identification of hospital-acquired ARF. METHODS: A panel (nephrologist, internist, clinical pharmacy specialist, pharmacoepidemiologist, database analyst) defined 3 algorithms based on existing literature and available automated data: 1) 50% increase of serum creatinine (SCr) within 3 days; 2) 50% SCr decrease between peak and discharge; and 3) ICD9 584.xx and charge code for dialysis. Each algorithm was linked (temporally and proximately) to drug exposure (aminoglycosides, amphotericin, cyclosporine, tacrolimus, NSAIDs, or radiocontrast). Discharges with hospital days < 2 or ESRD (dialysis in first 3 days of admission) were excluded. Algorithms were applied to the laboratory and administrative databases of a large teaching hospital including discharges between July 1, 2001 and June 30, 2002 (n = 20,639 or 10,536 with nephrotoxic drugs). Senior nephrology fellows to verify the algorithms reviewed a random sample of positive screened discharges. A random 20% of these were re-reviewed to assess inter-rater reliability. RESULTS: The 3 algorithms found 725 unique discharges (incidence 6.9%) with ARF. Algorithm 1 identified 585 cases, 2 and 3 identified 264 and 72. Of these, a random sample of 99 charts (stratified by algorithm) was reviewed. Reviewers anonymously confirmed ARF in 87 (88%) of 99 charts. The flagged association with the specified nephrotoxin was confirmed in 48 (55%) cases. Positive predictive values for each algorithm were comparable, 88%, 93%, and 87% respectively, while estimated sensitivity (78%, 47%, 15%) and specificity (25%, 75%, 83%) varied. CONCLUSIONS: The algorithms offer an excellent tool for outcomes research. While each is valid in correctly identifying ARF, differences exist in sensitivity and specificity. Underlying study questions (e.g., desired ARF severity) should determine which algorithm is used.