Adverse drug events (ADE) account for more than 100,000 deaths in the U.S. every year. Further, an estimated 250,000 preventable ADEs that result in various degrees of harm to patients occur annually. ADEs are associated with increased mortality, morbidity, and prolonged hospitalizations. This costs the U.S. taxpayers about $3-4 billion a year.
If you think about this, it is equivalent to a 747 jetliner crashing with 325 passengers aboard every day! If that happens once, it will be on the news for months. And yet we don’t hear about ADEs every day.
The timing of ADEs
According to the Agency for Healthcare Research and Quality (AHRQ), 49% of ADEs occur at the time of ordering the medications, 11% at the time of transcribing a physician order by another provider, 14% during the dispensing process, and 26% during administration of the ordered medications.
Electronic Medical Record (EMR) alerts are designed to fire during all these phases. The biggest bang for the buck is to identify and prevent ADEs at the time of ordering by the provider. While ordering a combination of drugs and fluids, the provider may not be well-versed with all the nuances of stability and compatibility, in which case an auto-generated pharmacy consult could help the pharmacy to combine and dose the drugs—relieving the physician from dealing with a bunch of annoying alerts.
Sometimes, nurses take orders from physicians over the phone by handwriting them on a piece of paper and then entering them online. This can result in transcription errors. Physicians are encouraged to write their own orders online to avoid this type of error. Errors can also occur if the wrong medications are administered to the wrong patients. But with the advent of positive patient identification using barcoded armbands, this error has been minimized.
Have EMR systems helped?
EMR systems were believed to reduce ADEs considerably, as they eliminated handwritten notes and orders. While that is true, unfortunately, EMRs also generate a ton of alerts and precaution messages to physicians while they interact with the system. On average, one alert fires for every order entered. Approximately 10 orders are written for a patient per day. So a 1,000-bed hospital generates around 10,000 alerts per day!
Because of this excessive number of alerts, providers have become overwhelmed and unresponsive to alerts, resulting in a condition called alert fatigue. This can lead to huge patient safety issues.
We hear a lot of complaints from providers that the majority of the alerts they receive are irrelevant and thus contribute to alert fatigue. Due to alert fatigue and the general assumption by providers that most alerts are likely to be irrelevant, some 90% of alerts are overridden. That means providers are dismissing some alerts that are relevant and could have a meaningful impact on patient safety.
The effectiveness of alert engines
Let us look at a simple way to evaluate the domain of alerts.
In the figure above, the circle shows the number of alerts fired. Five closed dots represent the relevant alerts and three open dots represent the irrelevant alerts. The open dots outside the circle depict the irrelevant alerts that are appropriately neglected by the alert engine. The closed dots outside the circle are relevant alerts that are missed by the alert engine.
Alert engines can be evaluated using the four measures as illustrated above:
- Precision or positive predictive value is the fraction of all relevant alerts that fired among all alerts fired.
- Negative predictive value (NPV) is the fraction of all irrelevant alerts that did not fire among all alerts that did not fire.
- Recall or sensitivity is the fraction of all relevant alerts fired among all relevant alerts in the universe.
- Specificity is the fraction of all irrelevant alerts that did not fire among all irrelevant alerts in the universe.
Let us move one irrelevant alert out of the alerts fired.
This improved the Precision, NPV, and Specificity. But no change in the Sensitivity.
Now let us move a relevant alert from the universe that is missed by the alert engine into the circle to be picked up by the alert engine.
This improved the Sensitivity for the first time as well as t Precision and NPV. But no change to Specificity. This indicates that alert management and optimization efforts should be targeted not only to eliminate irrelevant alerts, but also to uncover and include relevant alerts that are missed by the alert engine. A recent USA Today article discussing EMR systems noted that they fail to warn clinicians of approximately 13% of fatal errors.
Alert engine improvements on the horizon
Studies show that 45% of ADEs occur as a result of drug selection and dosing errors related to patients’ abnormal laboratory results. Out of that, a third are due to excessive dosing for patients with abnormal kidney and liver functions. Providers rely on their clinical knowledge to constantly monitor lab results and make modifications to a patient’s medications based on those results. Unfortunately, it is difficult for clinicians to remember all the precautions to be taken for the drugs that are affected by abnormal lab results. This could result in suboptimal treatment, or in some cases, harm to the patient.
The Clinical Drug Information business unit of Wolters Kluwer is working on a lab-driven clinical decision support system that would scan patients’ abnormal lab results and other elements of patients’ profiles and compare them against the drug profile to provide warnings and recommendations for appropriate drug choices and dosages. This will help tremendously in improving the number of relevant alerts that are likely to be missed by the alert engine.
Another strategy to help improve relevance of alerts is context enhancement. Physicians override alerts based on a number of factors, including alerts that are:
- less critical
- related to dosage forms/schedule/frequency
- not actionable
- not relevant to the current patient
- related to repeat orders
- not going to result in immediate harm
A system that is cognizant of these factors will be able to modulate the alerts.
Enhanced contextual alerting will also take into consideration potential interactions from dosages of drug pairs, age of the patient, and renal function to either fire or suppress the alerts. For example:
- Amlodipine and simvastatin interact with each other. But if the simvastatin dosage is less than 20 mg/day, the interactions are harmless.
- Severity of FluMist® and ASA interactions are age dependent.
- Severity of enalapril and spironolactone interactions are dependent on renal function and diabetes mellitus.
In order to receive relevant alerts, these specifics need to be factored into the alert logic engine.
When I talk to physician colleagues, they express a lack of trust in the system, since the majority of the alerts that fire are irrelevant when considering a specific patient’s personal profile. The only way to regain that trust is to make the system more intelligent by factoring in elements of a patient’s profile, including demographics, lab results, and other diagnoses. That will enable more appropriate alerting with a higher degree of Precision, Recall, Specificity, and NPV.
Wolters Kluwer is exploring other advanced clinical decision support enhancements, like predictive analytics and personalized medicine tools incorporating pharmacogenomics, to support the next generation of alert management.
The above figure summarizes the direction Wolters Kluwer is headed to optimize the alert management. Every patient’s unique factors are identified, based on the available patient’s profile, disease status, medication specifics, laboratory data, and known genetics, and then combined with filter settings and advanced rules logic. Alerts are then personalized to that patient, so they make clinical sense to the provider and helps earn back his or her confidence. That renewed confidence in EMR-generated alerts results in a behavior change of providers responding to those alerts favorably and using the valuable information and precautions to provide optimal care to enhance patient safety.