The Perils of Alert Fatigue

The authors of Alert Fatigue analyze and define the term as "...how busy workers become desensitized to safety alerts, and as a result ignore or fail to respond appropriately to such warnings."

They conclude:

1. Alerts are only modestly effective at best.
2. Alert fatigue is common.
3. Alert fatigue increases with growing exposure to alerts and heavier use of CPOE systems.

One of the potential great benefits of the electronic health record is clinical decision support, where artificial intelligence augments the baseline intelligence of the clinician.  The ultimate goal of CDS is to help the clinician 1- make accurate diagnoses', 2- use up to date treatment protocols, and 3- receive warnings about potential diagnostic and therapeutic pitfalls.

Troubles with intelligence (either artificial or natural) rise exponentially with the complexity of any issue.  For example, when one consults the PDR, every drug has hundreds of side effects and warning.  Ordinarily, the human brain filters these risks- many of which are hypothetical-and creates a mental hierarchy of potential dangers, on which action can be based.  However, when a computer performs the task of risk-retrieval, things become much more murky, since filters to prioritize, based on clinically relevant evidence, are pretty much absent, or if present, inapplicable.

Indeed most side effect warnings are inserted because pharma legal teams insert reference to every case report, not matter how obscure.  Otherwise- such non-medical counsels might argue- how could there be good defense against attribution of negligence, should a second atypical case pop up?  Putting such warnings into the adverse effect section of the ubiquitous PDR provides the possibility of escaping a lawsuit on the basis of physician or patient contributory negligence.  That is, the pharma defense team might argue, "they (the docs) did not (even) read the warnings!"

Attending to every one of the countless case-report engendered warnings would effectively take most medications- not to mention the clinicians- out of practical consideration.  On the other hand, some alerts are very serious and represent preventable errors that can potentially be avoided.  Balancing this voluminous input is hard on the human mind, and pretty much- without weighted prioritization- impossible for a machine.  Hence, the machine creates noise- repeated alerts- that the mind must deal with, and clinicians, getting inundated with endless warnings and signals, start ignoring them, even perhaps overriding their demands.

Human-factors come into play.  Protective desensitization to incoming stimuli is a necessary component of efficient cognitive activity.  Who among us has not become resistant to the noisy environment of an ED, just so we could get our baseline thinking in order?  The same can be said of a parent, working at home, with  chattering children running about.  Earplugs?  Well, sometimes.  But for the most part, many have learned to shut down the "noisy" input by "throwing an internal switch", one that no longer hears the kids screaming.  Repeated alerts, as we all have experienced, soon become just background noise.  Hopefully, thought, if the house is burning, the sound will be loud enough, different enough, for a red flag to arise.

This mirrors how an EHR vendor gets trapped into providing endless alerts.  For an EHR company, it becomes a business matter instead of a safety matter, created by potential legal liability.  If the physician doesn't read- and act- on a warning, the defense team might assert, it is not the EHR vendor's fault.  At bottom, these issues become matters of interpretation regarding what was serious and what was not serious when a legal case- or class-action suit- is initiated.  Vendors, understandably regarding their own interests, tend to be ultra-conservative, and this is especially true for CPOE of medications.

The authors suggest:

1. Increase alert specificity
2. Tier alerts according to severity
3. Make only high-level-severe- alerts interruptive

Even with such suggestions implemented, the problems are complicated since most enterprise EHR systems usually provide only globally effective solutions.  By nature, hybrid solutions do not achieve usability in specifically defined care-settings or circumstances.  The example we know best of a specific care-setting in which enterprise systems have failed to meet usability standards is the ED.  In other works, an alert that makes sense for one provider in one care-setting i.e. the ICU may not be appropriate elsewhere, i.e. the ED, nor would its intent and value necessarily be appreciated.

Some vendors have unfortunately left the job of tailoring alerts to the clinicians working at a client facility.  Dumping all responsibility on the docs is truly unfortunate, not only because it means considerable extra work and expense for the clinician staff but also because it nominally absolves the vendor of appropriate involvement in the effectiveness and up to date accuracy of the product.  The ability to do some local customization has definite value, but a balance between what the vendor inputs, and what the clinicians input, is necessary for a system to be effective, safe, and appreciated by all.

Our culture contains multiple myths and stories such as "The Boy Who Cried Wolf", and Chicken Little with the sky is falling.  Alert fatigue is a real entity and will need careful analysis to allow the positive clinical decision support to be a facilitator toward better clinical care and outcomes.