Health information technology (HIT) has been promoted as a method to improve health care delivery though coordination of care, workflow efficiency, clinical decision support and population health management. Informaticists could demonstrate their value within health care organizations by using HIT to advance each of these four domains within existing electronic medical record (EMR) platforms. I recently published a post about optimizing EMR decision support. Today, I will focus on optimizing HIT-clinical decision support through a generalizable model for implementing passive and active (interruptive, or “pop-up”) EMR alerts.
A 2017 review of computerized provider order entry and clinical decision support systems in intensive care units found meaningful reductions in medication prescribing errors and ICU mortality rates, but no change in length-of-stay or hospital mortality. In a 2013 BMJ meta-analysis of computerized clinical decision support, providing advice during electronic charting or ordering was associated with a reduced adherence to the HIT-generated recommendations. The authors suggest that this finding may be due to alert fatigue as most HIT-generated alerts are very sensitive, but not very specific. How might an organization use this information to improve the decision support delivered through their EMR systems?
Decision support is a Stage 2 Meaningful Use core measure for Eligible Professionals. The US Office of the National Coordinator for Health Information Technology states that clinical decision support “provides clinicians, staff, patients or other individuals with knowledge and person-specific information, intelligently filtered or presented at appropriate times, to enhance health care.” The ONC includes computerized alerts and reminders, clinical guidelines, condition-specific order sets, focused patient data reports and summaries, documentation templates, diagnostic support and contextually relevant reference information under as elements of clinical decision support. The organization has published a reference workflow taxonomy suggesting when different types of decision support may be considered. Provider-focused alerts are considered feasible in the following situations:
Outpatient encounter:
- During face-to-face encounter: Interviewing patients, ordering tests or prescribing medications, counseling patients
- Outside face-to-face encounter: Ordering studies or referrals, ordering or changing medications
Emergency room encounter: Interviewing patients, reviewing data from referring provider, ordering tests or prescribing medications
Outpatient surgical procedure:
- Pre-op: Pre-medication ordering
- Operating room: Documenting operative report, entering orders
- Post-op: Ordering tests or medications, discharge from recovery area
Inpatient hospitalization: Reviewing information, ordering tests or prescribing medications, hand-off or discharge
These four workflows suggest alerts during chart review, while interviewing a patient, when ordering tests or medications, during clincal documentation and preparing for a care transition. Chart review and patient interviews seldom occur with the clinician interacting with both the patient and the EMR, limiting an alert’s effectiveness. If providers reliably reviewed charts before speaking with patients, chart review decision support to guide follow-up testing for historical inconclusive test results, patient-specific screening tests and disease surveillance may be beneficial. Documentation alerts could prompt clinicians to document specific elements, but natural language processing reviewing note content prior to signature might be more effective in helping users adhere to documentation requirements.
Active alerts may be more appropriate for specific ordering scenarios and alerting users about specific actions before moving the patient to the next care setting. In general, firing alerts when an order is selected but before it is configured reduces the re-work a user might have to do based on the alert content. Selecting a transfer or discharge order may indicate the last moment before a provider will think about the patient’s outstanding tasks. Interrupting the user at this point may be appropriate for specific urgent organization or clinical requirements.
EMR provider-focused alert-based decision support could be generalized with the following questions:
- What workflow will we suggest to help target decision support at the right user at the right time?
A hospital admission workflow might be:
A. Prior to seeing the patient, review the chart to identify gaps in care (passive alerts).
B. See the patient (no alerts). Savvy EMR users might update the patient’s medication list, problem list, patient’s primary care provider and other data elements. Users who would prefer not to use the EMR in front of patients may need to spend more time completing their documentation after the visit.
C. After seeing the patient, update patient’s encounter diagnoses and problem list (enables EMR to trigger disease-specific alerts in the future).
D. Complete the patient’s History & Physical with or without voice recognition (decision support around a broader differential diagnosis or clinical documentation improvement).
E. Open the EMR’s Order Entry screen for admission orders (active alerts about missing medication allergies and patient-specific clinical deterioration).
F. Select specific orders for configuration and signature (active alerts for duplicate tests, duplicate drug, drug-allergy, drug-drug and drug-disease interactions).
G. Sign admission orders (active alerts for missing level of care order, advance care planning, disease-specific alerts, admission medication reconciliation and justification for two-midnight rule).
This specific workflow enables decision support based on when the clinician has access to the EMR and/or the patient to gather additional information. Other workflows might highlight other moments to provide decision support.
- Who should receive the alert? How should repeated alerts be handled?
After determining where the alert should fire, limiting the alert to specific roles or individuals will reduce the number of unnecessary alerts. Limiting the specific alert to those users who can perform the task or place the order to address the alert will help all users understand that the alerts they do see are specifically intended for them.
An alert that fires more than once for a provider caring for the same patient within a short time frame (e.g., three months in the outpatient setting, same hospitalization in the inpatient setting) may need to be deployed differently based on the alert’s intent. An alert addressing the same drug-drug interaction that was present 24 hours ago without any patient adverse effects could be considered for suppression. A patient qualifying for a sepsis evaluation 24 hours after an initial qualifying event should trigger a new alert to the care team.
Some users may repeatedly ignore passive alerts. Linking an active alert to these unresolved alerts before a patient leaves a specific care setting (e.g., signing a transfer or discharge order) reinforces the importance of passive alerts without forcing organizational leaders to demand more pop-up alerts.
- Would additional information help tailor an alert?
For most clinical scenarios with risks and potential benefits, knowing a patient’s treatment preferences would help inform what decision support would be most appropriate. A patient under hospice care may consider mechanical ventilation differently than a patient who expects to live another 20 years. Recording patient preferences in a discrete way can help further tailor alerts to the intended patient population.
- How will we know if an alert is working?
This question suggests defining measures of success initially and long-term. When an alert is deployed initially, users may ignore the alert. Pre-specifying when and how an alert might be updated will reduce the risk of hastily-designed updates or allowing the alert and its associated organizational objective become background noise within the provider’s EMR experience. End-user behavior should be reviewed regularly to update the alert. Once users seem to be performing the intended behavior as expected, consider de-escalating the alert to a passive alert to make room for new organizational initiatives (e.g., dropping venous thromboembolism prophylaxis as an active alert for an active sepsis alert instead).
If the alert is important enough to interrupt providers, it should be important enough to track against its intended objective. Demonstrating the link between the alert and the organization goal can help keep the clinical and EMR leadership teams focused on how active alerts might help employees and other health care team members perform unfamiliar tasks with higher levels of reliability than might be expected with education or post-hoc audit and feedback.
Although there are many different forms of clinical decision support, clinical end-users seem to hate interruptive alerts the most. Using a framework based on standard workflows, passive alerts linked to active alerts and explicit connections to organizational objectives may help simultaneously improve provider satisfaction and increase care delivery quality.