In contrast to clinical reference material, such as continuously updated online journals and medical texts, computerized clinical decision support directly assists the clinician in making decisions about a specific patient. Decision support does not need to be sophisticated to have significant impact. For example, simple dose-range checking for medications, drug-drug interaction checking, and drug-allergy checking are conceptually straightforward but can catch a critical source of human error that no amount of training or personal vigilance can entirely eliminate. Slightly more advanced are systems that analyze clinical data (such as calculating a creatinine clearance) and present guidance based on those data. More sophisticated systems look for trends in values, such as the rate of fall of the hematocrit or the rising body weight of an intensive care unit (ICU) patient who is accumulating extracellular fluid, where an absolute number may not be notable, but an alert to the trend may be important and prompt action.
Decision support is challenging to implement and maintain. The most vexing problem is “alert fatigue.” Studies within and outside health care show that the benefit of an alert, such as a pop-up interaction in a software system, is rapidly extinguished if the alert becomes a routine part of using the system. A familiar clinical example is the minimal attention paid to most audible alerts produced by cardiac telemetry systems. If a clinical decision support in a CPOE system provides an “alert” to the drug-drug interaction of two medications routinely used together safely, such as enoxaparin and warfarin, in the same way as to unfamiliar but dangerous interactions, such as theophylline and fluoroquinolones, clinicians become desensitized to the alerts and dismiss critically important guidance when it does appear. Alert fatigue is a fact of human cognition and cannot be eliminated through training, education, or self-discipline. The best clinical systems offer fine-grained tuning of the system's behavior, such as altering the system's response by drug and provider specialty, and offer a range of interruptive and noninterruptive support mechanisms. The price of this flexibility is the institutional effort required to design and maintain the system. However, even these measures have yet to demonstrate consistent improvement in the effectiveness of alerts.
The most complex decision support systems attempt to aid clinical diagnosis. The application of artificial intelligence to medicine has a long history; however, most diagnostic expert systems have been stand-alone, requiring effort by the clinician outside of their normal workflow, and have thus had limited clinical implementation. Examples of clinical diagnostic systems directly imbedded in an electronic health record are few but are an area of increasing research and commercial interest (Table e3–2).
Table e3–2.Functional classes and examples of clinical decision support systems. ||Download (.pdf) Table e3–2. Functional classes and examples of clinical decision support systems.
|Class ||Function ||Examples |
|Feedback ||Provide feedback by responding to an action taken by the clinician or to new data entered into ...|