Diminishing Noise at the Point-of-Care: 4 Steps to Holistically Battling Alert Fatigue

Better information for better care delivery — it's one of the primary goals of the health IT movement. As federal initiatives continue to push for improved decision-making and increased uptake of evidence-based medicine, the introduction of clinical decision support at the point of care has been identified as a critical step.

And rightly so. Technology that arms clinicians with the latest industry evidence when they need it most helps minimize the potential for mistakes and adverse events and ensures the most cost-efficient care delivery based on industry best practices.

Few in the industry would dispute the potential of point-of-care CDS to elevate care. The problem to date has been positioning technology to deliver the right balance of "alerts" that addresses not only the needs of clinicians but also the safety the patients.

Studies reveal that when clinicians are exposed too frequently to alerts they feel are irrelevant to their patients or workflow, they will become desensitized and begin to ignore potentially relevant alerts. Simply put, too much "noise" at the point of care can drown out the real opportunities to avoid harm and enhance patient care.

While alert fatigue can be problematic across many areas of EMR workflow, it is particularly prevalent with medication orders. In fact, medication alerts are so common that they have created a situation where "systems and the computers that are supposed to make physicians' lives better are actually torturing them," according to the author of a 2009 study of nearly 3,000 prescribers in three states. The study found physicians ignored alerts more than 90% of the time.[i]  

Alert fatigue remains one of the greatest hindrances to optimal use of clinical decision support technology at the point of care, but the outlook is improving and will continue to do so as vendors advance functionality to address the challenges. The best approach to combating alert fatigue so that CDS technology can be leveraged to its full potential is a multi-faceted one that combines flexible IT infrastructures working in tandem with appropriate governance processes.

Consider these four critical steps:

1. Tiered alerts
Studies reveal that tiering alerts based on severity can increase compliance rates. A 2009 study examined the impact tiering drug-drug interactions by severity level had on compliance and found that compliance was higher when alerts were tiered in this manner.[ii]  Specifically, 100 percent of the most severe alerts were accepted at the tiered site, compared to only 34 percent at the non-tiered site.  Moderately severe alerts were accepted less frequently in both cases, but they were still accepted more often at the tiered site (29 percent versus 10 percent). Researchers therefore noted that failure to tier resulted in ""substantially less recommended provider behavior."

2. Contextual alerts
Making alerts smarter via the context in which they are delivered makes the decision support provided to a clinician much more relevant. Some alerting systems are already designed to provide functionality for contextual alerting content that considers patient profile data such as symptoms, age, weight and gender.

Context for drug-drug interactions is important. For instance, consider drugs that might raise a patient's potassium level. If a patient's potassium is low to begin with, using a particular drug might not be of much concern. In the case of combining two drugs like spironolactone and lisinopril, an increase in potassium could produce hyperkalemia, possibly with cardiac arrhythmias or arrest.

A recent Journal of the American Medical Informatics Association paper found that a machine learning approach informed by expert feature selection and expert classifications could prove to be an effective way to create a contextual model.[iii]

The study identified features by having experts review a small number of alerts and verbalize the data they reviewed as part of their assessment.  Expert opinions were also provided on the importance of anemia alerts for 118 cases. This training data set was fed into a machine learning system, and the results were tested on 82 new cases. The system had a precision of 87 percent for low level alerts, meaning that 87 percent of the alerts the computer classified as low level had also been classified by the experts as low level.  This paper suggests that expert feature selection and machine learning could be used to generate contextual models for alerts.

3. Customization by organization, department and clinician
Applications that support user control and customization of alerts on a departmental or individual basis can go a long way toward addressing alert fatigue.  Alerts that are relevant to one group of clinicians may have little meaning to another.

For example, the combination of the antibiotic linezolid and the antidepressant fluoxetine can cause central nervous system toxicity.  Infectious disease specialists tend to be very familiar with this antibiotic and this possible interaction, so they may consider such an alert to be noise.  On the other hand, general internists or family physicians may not be aware of the issue, so an alert could prove to be very important for patient safety.

User control is just one more way of reducing noise at the point of care and allowing clinicians to maintain some autonomy over the process. This kind of functionality can be readily embedded in the clinician's workflow with a checkbox saying, "Don't show me this alert again" or "Don't show me this alert again for this patient."

4. Effective governance
Governance over processes can come in a number of forms, but at a foundational level, healthcare organizations need to identify a committee to govern and manage CDS content and applications. In regards to alerts, this committee would make decisions about how filters might be set for tiered and contextual alerting and user controls.

Before making those kinds of decisions, it is important to have data to back up the decision-making process. An analysis should be conducted to determine the types of alerts that appear most often, as well as which alerts are being overridden and which are not.

This kind of monitoring is effective for initial decisions as well as ongoing management.  Monitoring override rates on an ongoing basis and utilizing the resultant data to reclassify the severity of alerts can help healthcare organizations fully leverage the potential of their CDS.

Conclusion
Alert fatigue remains a significant barrier to healthcare organizations adopting and realizing the full potential of CDS. Vendors are making significant strides to deliver decision support that is smarter and more relevant to an individual clinician's workflow, but more must be done in order for the industry to fully leverage the potential of CDS to impact quality and safety. Going forward, technology must provide functionality to address tiered and contextual alerting as well as user control to deliver the kind of smart alerts that will minimize the potential for alert fatigue.