Identifying adverse events by chart review, as in the Harvard Medical Practice Study and other major epidemiologic studies in patient safety, is a time- and labor-intensive process. Most institutions could not afford to make such investments of resources to identify adverse events except in the setting of a funded research study. To address this issue, the authors applied a commercially available search engine to scan for scanning discharge summaries to identify adverse events. Using a set of cases with known adverse events, the investigators developed rules based on combinations of terms associated with adverse events. The sensitivity of this approach using the particular combinations of terms examined by the authors was only 23%, but the specificity was 92%. Thus, this approach may provide an efficient means of flagging cases for internal review. In other words, while this natural language processing approach misses cases, the cases it does identify have a high probability of representing true adverse events. Running this search on a regular basis would yield cases likely to reflect quality or safety problems without requiring the same investment of resources as would screening for cases through manual chart review.