Reforms recommended for CPOE alert systems

Mark Siska, medication clinical systems manager of pharmacy services at the Mayo Clinic in Rochester, Minn., and Scott McCreadie, Pharm.D., strategic projects coordinator for the University of Michigan Health System, addressed these issues in their presentation at the 2006 Annual HIMSS (Healthcare Information & Management Systems Society) Conference in San Diego.

The CPOE system at the University of Michigan Health System triggered alerts for 90% of new orders in January 2003. McCreadie emphasized, "That's almost every single new order." Several recently published studies have found that most CPOE-generated alerts are overridden-with an override rate as high as 91% in one study. Reasons for overrides include the patient has tolerated the drug in question in the past and the patient is currently taking (and tolerating) the drug in question.

Siska and McCreadie made several recommendations to improve these systems. For starters, CPOE databases need to be purged of false positive alerts. At the University of Michigan, a systematic effort to reduce alert noise by careful tuning to filter out false positives decreased the number of alerts by 40%.

Knowledge bases should be customizable by users. Pharmacists, who have been familiar with alerts for decades, should have a role in revising the alert systems. According to Siska, "Knowledge-base vendors' editorial review policies adopt an extremely conservative approach when defining an alert, emphasizing breadth, and are highly inclusive, with low thresholds for alerting."

Alerts should be stratified according to different levels of severity, with screen displays designed to reflect the severity. An alert for possible abdominal pain from an antibiotic being prescribed for a patient facing death from anthrax shouldn't receive the same attention as an alert warning about a lethal interaction between the antibiotic and a drug this patient is currently taking. Low-level alerts should be noninterruptive. Intermediate-level alerts would require justification. High-level alerts would block the order.

Alert systems need to be flexible enough to accommodate different disease states, different patient histories, and different users. A physician prescribing an antibiotic for an inpatient near death doesn't have time to receive an alert that the antibiotic many cause dizziness. On the other hand, for an outpatient with an infected thumb who happens to be an airline pilot, dizziness could be a problem necessitating an alternative treatment.

The integration of medication alerting systems with other data sources could provide that kind of flexibility. For example, a medication alerting system could interact with patient lab data. McCreadie pointed out that many drugs are affected by blood potassium levels. If the medication ordering system were integrated with patient lab data, an alert could be triggered when a patient with high potassium levels is about to be prescribed a drug rendered ineffective by high potassium. Of course, the prescriber would never see this alert when the patient's potassium is normal. Thus, we see how interaction between databases could both shorten the time it takes to write an order and increase patient safety.

McCreadie and Siska also recommended that adopters of CPOE do research to assess how well their CPOE systems are working and publish their results so that they would be available to both potential CPOE adopters and current CPOE users tweaking their systems. Research is needed to recognize potential shortcomings of current systems, assess the consistency of existing alert databases, identify the most effective means of communication between providers, and identify best practices.

McCreadie summarized by saying, "The most important message is that physician medication alerts in a CPOE environment must be approached differently than drug alerts in pharmacy systems. Simply turning on alerts from a drug reference database will not work. Physician alerts must be highly tuned to reduce false positives." Furthermore, he said, the best systems are those that interact with hospital databases to suppress alerts that are not relevant to the patient in question.

THE AUTHOR is a writer based in San Diego.