Electronic medical record decision support is intended to increase the likelihood that the most appropriate care will be delivered every time. Information technology teams managing electronic medical record software face competing demands to successfully deploy decision support. Health system leaders expect health information technology to enforce adherence to broadly accepted care pathways, adopt emerging science and alert users about potential diagnostic or treatment errors. Front-end clinical users generally find electronically-generated guidance to be unnecessary and cumbersome. Patients expect health systems to identify their preferences to tailor their encounters accordingly. Developing a decision support paradigm that aligns these different stakeholder interests may simultaneously improve clinical care delivery without overwhelming the analysts supporting the electronic medical record.
Health system decision support expectations
Compared to paper-based medical records, electronic medical records restrict what clinicians can do. Medication orders in an electronic medical record require a dose, route and frequency before being processed. Some documentation templates require specific fields to be completed before electronic signature. Electronic systems can present users with warnings that must be addressed before completing a workflow. As health systems become more sophisticated, they will implement these “forcing functions” to drive clinicians toward specific care pathways to achieve specific objectives (e.g., ACE-inhibitor prescription or document a contraindication to ACE-inhibitors for patients discharged with systolic congestive heart failure). Pathways linked to publicly-reported measures may take precedence over pathways that meet other objectives. Each clinical service line will argue for their specific care pathways without regard to the overall impact on clinical end-users.
Beyond clinical consensus, new research findings push decision support stewards to incorporate emerging content into existing workflows. Clinical trials use detailed inclusion and exclusion criteria to help identify specific cohorts that might benefit from a particular diagnostic or therapeutic intervention. Leading health systems would like to implement these findings into their care processes as an indicator of superior care. Rather than forcing clinicians to recall specific indications for a particular intervention, health system leaders will expect electronic medical record systems to ascertain what patients meet the relevant criteria to be considered for a novel intervention. In the most extreme case, interventions based on genetic markers or other elements of personalized medicine will require implementation within an electronic medical record.
Safe clinical care is increasingly defined by how health information technology is employed to ensure patients do not incur additional risks during a clinical encounter. Computerized provider order entry, scanning bar-coded therapeutic agents (medications, blood products) and pre-procedure checklists make health care delivery safer. As hospitals and health systems are under increasing pressure to prevent iatrogenic injury, teams managing electronic medical records will be asked to refine their own processes and integrate other health information technology to support safer care delivery.
Clinical end-user decision support expectations
Clinical end-users struggle with the number of suggestions, warnings, and workflow interruptions that confront them when using the electronic medical record. Some interruptions are redundant, others have little value, and still others are not relevant to the specific end-user presented with the decision support. When overwhelmed with these interruptions (i.e., alert fatigue), users override all alerts, reducing the likelihood that relevant decision support will be reviewed and acted upon. In one emergency room study, the alert override rate was over 95%.
Like most of us who use technology, clinical end-users tend to stick with workflows that help them achieve their objectives, even if new functionality exists that allows them to complete their work in a more safe and efficient manner (Have you switched from credit cards to mobile payments? If you haven’t, you aren’t alone.). Older users employ workflows based on a user interface that was current at the time of initial training. New users learn the most modern version of those same workflows. Older users will watch newer users complete a task, express some appreciation, and then resume their legacy way of completing the task. In my experience, the only instance when users switched workflows on their own volition was using mobile devices to access to electronic medical records with voice-to-text technology. And even then, the technology was adopted with variable penetration among our provider user base. In all other instances, older workflows had to be disabled to migrate users to newer workflows. The heterogeneity in how clinical tasks are performed limits how decision support can be targeted so all users who need to see a specific decision support tool actually see it.
Decision Support Framework
Jerome Osheroff has been a leader in this field. The framework below is my interpretation of his work defining the five rights of clinical decision support (right information to the right people in the right formats through the right channels at the right times) applied to an inpatient electronic medical record with both academic and community facilities:
- Passive decision support
- Chart review
- By data type (operative reports, chemistry results)
- By specialty (cardiology, anesthesiology)
- By disease or clinical state (diabetes, total parenteral nutrition)
- Specialty order favorites
- Providing contextual information as part of the ordering process (e.g., displaying the patient’s last three glomerular filtration rates when prescribing nephrotoxic antibiotics)
- Order sets triggered by a test result or diagnosis entered into the patient’s encounter diagnosis list or problem list
- Specialty documentation favorites
- Templates including prompts for required elements (estimated blood loss, implants for operative reports)
- Chart review
- Interruptive decision support (pop-up alerts)
- Within the patient’s record: opening a patient’s chart, proposing or signing an order (e.g., drug-allergy, drug-disease [hepatotoxic drugs for patients with poor liver function], drug-drug interactions, or duplicate test alerts)
- Outside the patient’s record: a notification to a pager or monitored work queue
*Telephone dictation and other documentation methods that import data into the electronic medical record through interfaces reduce the effectiveness of documentation-based decision support.
Patient decision support expectations
Patients spend most of their time outside the health care system. Although some patients may record information regularly for clinicians to review during health care encounters, others may not see the value in tracking health-related information after leaving a health care facility. Some patients will view health care visits as the only avenue to access tests and treatments. To meet a patient’s expectation that health care systems incorporate all relevant data when agreeing to pursue a patient-suggested intervention or suggesting their own intervention (e.g., ordering a test, making a referral, starting or intensifying a treatment), electronic medical record decision support should support clinicians in:
- Obtaining information from other health care settings (possibly mediated by the patient),
- Summarizing patient-generated health data,
- Eliciting patient preferences (subject to change over time),
- Retrieving clinical evidence that might inform a decision to pursue an intervention, and
- Communicating the risks and benefits of a particular intervention to the patient and caregiver to help make a shared decision.
Optimizing Decision Support for Today
Effective decision support relies on stratifying users into as many distinct specialty- and discipline-specific groups as the electronic medical record team can manage. Each subgroup can then have their own chart review views, order preferences and documentation template preferences. Interruptive alerts can be filtered to only those subgroups can act on the information presented. As the users mature in their experience with the electronic medical record, the smaller groups are more likely to come to consensus to support changes to their screens rather than a larger group of unrelated users.
Order sets are the primary vehicle to drive ordering behavior. Order sets that flex based on a patient’s age, gender and specific clinical parameters are more likely to be accepted than a single set of orders displayed to users treating different types of patients. Prior test result information and details about when an order might be completed may help guide users to determine if the order will add value to the patient’s care. Hyperlinks can be included within order sets to provide end-users with information around risks and benefits of different interventions as well as information about what patient preferences might drive different ordering choices. Finally, order sets with tasks distributed across different team members over time should include some visual cue to help the user determine where the patient is positioned in an expected workflow. The information can help facilitate communication among team members to help advance the patient’s care appropriately.
Users find interruptive alerts the most irritating type of decision support. Some workflows might benefit from a checklist to show the user what elements are missing to complete a specific process (e.g., admission, discharge). If a user completes a specific element expected at the end of the process (e.g., signing a discharge order), the user could then be alerted to complete the outstanding tasks being allowed to sign off on the final element.
Alerts that interrupt a user in the ordering process can be especially challenging. Electronic workflows can be executed much faster than paper-based ones, increasing the need to stop incorrect orders before they are signed. These types of alerts could be stratified into initial and subsequent alerts. Initial alerts may highlight a drug-drug interaction for a specific user. If the user determines the benefits outweigh the costs, that decision should then prevent subsequent alerts of the same type from appearing for that user within that patient’s record. This approach should reduce the total number of alerts a user views without jeopardizing patient safety.
For decision support reminding users to complete a task that does not interfere with clinical orders or documentation (e.g., co-sign orders, overdue operative report), there is no reason to interrupt the user before leaving the encounter. One method to scale decision support across specialties and clinical disciplines would be to aggregate these reminders within a section on the patient record’s “home page.” The section would need a visual cue to remind users to check the section when tasks need to be reviewed and addressed. Periodic audits would need to be performed to help reinforce the need for clinicians to review these reminders without an alert.
Designing Decision Support for Tomorrow
As our health information technology tools improve, we should augment our decision support tools accordingly. In 2008, Sittig et al. identified 10 challenges for clinical decision support. We as a profession have made progress on some (summarize patient-level information, prioritize and filter recommendations to the user), but not others (use freetext to drive clinical decision support, combine recommendations for patients with co-morbidities). As clinicians become more comfortable with incorporating new modalities (mobility, voice) into their workflows and electronic medical records incorporate more usability learnings into their future releases, decision support could help improve how health care is delivered within traditional care settings in the following ways:
Chart review (before seeing the patient)
There are at least three different sources of information that might provide value before seeing the patient:
- Prior health care encounters – The easiest records to access are those within the same electronic medical record platform or within the same health information exchange. Diagnoses with associated complications, discharge summaries and laboratory information can provide the necessary context to tailor decision making for the current encounter. Surescripts has become the default method to verify medication fills. Patients may also share records from other facilities directly with providers.
- Patient preferences – preferences that impact how a clinical encounter is conducted include language; decision-making style; readiness to change specific health behaviors; willingness to make trade offs among effectiveness, availability, costs (immediate and downstream) and side effects; and tolerance for uncertainty. Even if the provider is meeting with the patient for the first time, the patient may have dealt with the symptom or condition over multiple health care encounters and could have preferences that will affect how decisions might be made during the encounter.
- Patient-generated health data – Blood pressure measurements, blood glucose results, sleep diaries and functional assessments are examples of data feeds that may overwhelm providers without tools to summarize this information quickly and accurately.
Tools that can summarize these data sources by symptom or disease will help clinicians understand and incorporate the information into the face-to-face patient encounter. Rather than rely on a problem list, decision support using natural language processing and a longitudinal health record could help clinicians understand the certainty of a particular diagnosis (clinician judgment, response to treatment, diagnostic test results) and its management over time. Understanding what conditions have the largest impact on the patient’s quality of life might help the provider better negotiate the visit agenda to start the encounter. The absence of specific information (e.g., gaps in expected care, treatment preferences) can help providers target those areas for further discussion with the patient.
Documentation (either in front of the patient or just after seeing the patient)
Clinicians who update medication allergies, medications and encounter diagnoses in real-time (or near real-time) will facilitate subsequent decision support. Family history may also be relevant for detecting inherited conditions.
After reviewing the patient’s medical record, interviewing the patient and performing a physical exam, voice-to-text technology is now good enough for a provider to complete their initial documentation (history & physical, consult) before placing orders. When the user authenticates the document, natural language processing algorithms could identify gaps in the documentation from a regulatory or payment perspective (e.g., no Review of Systems for a consult). The algorithms could also identify possible diagnoses matching the patient’s demographics and presentation.
This diagnosis identification allows for two different decision support interventions. First, discrepancies between the patient’s diagnoses included in the documentation and the discrete diagnosis fields (encounter diagnoses, problem list) could be reconciled. Clinical pathways and other decision support tools rely on accurate discrete diagnoses to be triggered. Second, users could update their documentation to include diagnoses identified through the decision support by either add information to support a different diagnostic strategy to consider a new diagnosis or enter details that why a particular diagnosis is less likely to be relevant in this case.
Documenting relevant findings and the rationale for next steps before entering orders may be the only way to interrupt actions that may not help or potentially hurt the patient. Providing decision support earlier in the diagnostic process is critical as clinicians will anchor on their first diagnostic impression and are more likely to reject later information suggesting an alternative diagnosis. AHRQ’s Patient Safety Network describes some other heuristics that might be challenged with more real-time decision support.
Ordering (after the clinical encounter)
Many electronic medical record system ordering modules have migrated to a “shopping cart” paradigm. Users are encouraged to enter all of their orders before clicking “Sign.” Ordering windows could be tailored to highlight orders or order sets that are relevant for this patient based on the pre-encounter information and current visit documentation. For those decisions that would benefit from understanding patient preferences, providers could order decision support modules that could provide language, numeracy and health-literacy calibrated information for patients and caregivers to review before the next clinical encounter. A patient’s insurance status may dictate that an intervention be performed in either the inpatient or outpatient setting. Decision support could help users guide their patients toward the most cost-effective option. As users added more orders, the software could dynamically update what orders might be recommended or removed. Some tests added to shopping cart might require additional decision support (computed tomography and magnetic resonance imaging tests) while conflicting orders may require reconciliation before signature.
Interruptive orders at the “Sign” moment should be reserved for orders that have a significant likelihood of harming the patient. If these orders can be identified during the order selection process prior to signature, the clinician should have the opportunity to address them before signature (e.g., “I am aware of the patient’s renal insufficiency and believe the benefits of [medication] outweigh the risks in this patient.”). This approach allows users to view the decision support as workflow-congruent instead of stopping users from completing their work.
After seeing a patient for the first time, the clinical user should have a sense of the patient’s medical history and current medical conditions. Daily rounding in the hospital or follow-up visits in clinic should be less complex than an initial visit. Clinicians might use mobile devices to review interim updates before walking into a patient’s hospital room or clinic exam room. This “just-in-time” data retrieval assumes all of the relevant information is viewable and actionable on a mobile device.
Chart review (before seeing the patient or during the initial phase of the clinical encounter)
Summarizing clinical data generated since the last face-to-face encounter should help orient the provider to the patient’s current condition. As signs and symptoms are classified into distinct diseases with additional information, decision support could highlight what additional testing and treatments might be helpful. Reviewing the disease trajectories of similar patients could be used to determine if the patient’s response to therapy is consistent with others diagnosed with the same condition or if the diagnosis should be revised.
Decision support could also be used to identify those interventions that might have the greatest impact on the patient’s quality of life. Estimating the value of a patient increasing their medication adherence to 80% for a subset of medications might help increase the odds the patient makes the necessary behavior changes to improve their health. This type of decision support would require an accurate identification of the patient’s preferences and some level of evidence that the intervention has value from the patient’s perspective. After having a conversation with the patient, the provider should be in a better position to determine what orders to place next.
Orders and documentation (after seeing the patient)
Decision support for orders and documentation for subsequent encounters should follow the same principles as seeing a patient for the first time. One additional function within documentation decision support might be to identify those findings or assessments that merit an update in today’s documentation based on prior documents, test results or changes in the patient’s condition.
Urgent requests outside the patient’s chart
Tasks that need to be completed sooner rather later (e.g., co-sign a level of care order to qualify for reimbursement) could be managed using one or more of the following tactics:
- Mark the patient’s record with a special indicator that could be viewed by a user who is logged into the electronic medical record, but not the patient’s chart (e.g., a patient list)
- Route the task through an escalating pathway outside the patient’s chart (message to an electronic medical record mailbox, secure text message to the provider, secure text message to provider’s supervisor). This tactic requires some understanding the members of the care team to reduce task routing errors.
- List the outstanding tasks in a work queue for subsequent triage in the event no user authorized to complete the task has logged into the electronic medical record or has access to their secure texting program
Health system leaders working within a clinical governance structure would have to determine what tasks meet this threshold to balance the interruption’s cost against the value to the organization.
Preparing for the next care setting
Hospital discharge planning is a process worthy of its own decision support. In addition to a checklist to make sure the necessary steps are completed, more advanced decision support could help identify those actions that have been suggested over the course of a hospital stay. Most clinical users will rely on history & physicals, consults, operative reports and discharge summaries to accurately capture a patient’s presentation and hospital course. Describing what findings were pursued and what diagnostic or treatment suggestions were not addressed is essential to help clinicians provide a seamless transition across care settings. Today, the work to determine the quality of a discharge summary is too time-intensive to be feasible. Applying currently available technology to this problem could allow real-time feedback to clinicians to improve the quality of their documentation.
The Office of the National Coordinator for Health Information Technology lists decision support as one of the reasons clinicians should adopt electronic medical records. Rather than require incorporating decision support within the electronic medical record, Medicare is moving toward electronic submission of quality measure performance for hospitals and providers, implying some type of decision support or quality improvement drives clinicians to achieve specific outcomes. Updating electronic medical record decision support based on health information technology advancements and clinician readiness to adopt new workflows will help provide the suggested value gains when fully adopting health information technology when caring for patients.