Information overload and missed test results in electronic health record-based settings.

cal one, especially because of the rapid proliferation of decision aids. However, reducing alert fatigue comes at the expense of a higher rate of false negatives (lower sensitivity). If the alert threshold were increased from 40% to 80%, this would increase its specificity even further, but such an approach would have limited clinical utility. The best way to determine the right threshold is to prospectively test the tool in real-world settings and examine how it changes physician treatment behavior and/or the outcomes and efficiency of patient care. What about the use of chest radiography as part of the computerized screening tool? The chest radiograph is considered the reference standard for the diagnosis of pneumonia, despite the fact that it is neither 100% sensitive, nor 100% specific. Pneumonia can be present in the absence of an acute radiographic infiltrate, particularly among elderly patients with dehydration at the time of presentation. In addition, substantial interobserver variation among ED physicians and even radiologists has been well documented for the interpretation of chest radiographic findings compatible with pneumonia. Notably, 59% of the false positives of the reported screening tool were due to incorrect reading of the imaging reports using natural language processing. If trained radiologists cannot agree, it comes as no surprise that the tool itself also has difficulty interpreting radiography reports. Perhaps a more basic need in the diagnosis of pneumonia is to standardize the interpretation and reporting of chest radiographic findings prior to attempting to integrate such findings into a computerized decision aid. This innovative and unique tool may foreshadow the future of medicine—a future in which vast amounts of data are synthesized by computers to improve physician diagnosis and treatment decisions. As others have discussed, the quality and accuracy of diagnosis is often overlooked in most efforts to improve quality and safety, which instead focus on the management of alreadydiagnosed problems. Although there is clearly a role for decision-support tools, prior to their wide adoption they will need to show an incremental benefit over physician judgment alone and demonstrate improvements in patient outcomes.