Personalized Nudges Aren’t Just for Patients; They’re for Clinicians Too

Therapeutic inertia (TI) is defined as the failure to escalate treatment when warranted. In a recent study where neurologists were exposed to simulated case scenarios of patients with multiple sclerosis (MS), it was found that over 70% of neurologists demonstrated TI during treatment decisions. This led to suboptimal decisions in approximately 20% of cases.¹

Interestingly, the research group was able to conduct a deeper analysis and identify certain physician characteristics that were associated with an increased likelihood of TI. These factors include aversion to ambiguity, level of expertise in MS care, and even country of practice.

The same group also developed a short educational intervention that leveraged what is known as a “traffic light system.” The traffic light system uses color associations to help clinicians identify high-risk patients (i.e. red signifies high-risk and escalates treatment while yellow encourages re-evaluation in 6 months). More details about the intervention can be found in their study. A randomized controlled trial proved its effectiveness in reducing the prevalence of TI during MS care. This is promising for improving clinical outcomes for MS patients.²

It is important to note that the traffic-light intervention was not delivered as a nudge in this study. Rather, it was delivered to study participants that were given case scenarios that represented real-world clinical situations. However, in a real-world clinical setting, it could be presented as a nudge by being delivered through electronic health records or another format during patient care.

Statins have been proven to significantly lower the risk of cardiovascular events and mortality. Several studies, however, have demonstrated their underutilization: about 50% of patients who would benefit from statins (according to national guidelines) are not prescribed them by physicians.³ On top of this, those who are prescribed statins more often than not receive a lower-than-optimal dose. Researchers at the Penn Nudge Unit, a behavioral design team embedded within the University of Pennsylvania Health System, sought to evaluate whether nudging could improve guideline-concordant statin prescribing.4

Their first study compared a control with two intervention arms: an active choice intervention with and without peer comparison performance feedback. Physicians in the first arm (just the active choice intervention) received an email that included a list of their patients who would benefit from statins but were not yet prescribed them. Other data in the email included relevant patient clinical characteristics and the national guidelines for statin prescribing. Here, the active choice nudge was asking the physician to review the list and select whether or not to prescribe a statin within 1 week.

In the peer comparison arm, physicians also received information about how their statin prescribing rates compared with those of other physicians at their institution. This nudge was based on what is known about social norms and their effects on behavior change. The active choice + peer comparison intervention arm demonstrated a significant increase in appropriate statin prescribing. 

In another study, the same research group identified clinician characteristics that were associated with statin prescribing. Specifically, being female, being a physician assistant, and having more years of experience were associated with a lower likelihood of prescribing statins.5

Why does personalization hold so much promise for the future of clinician nudging? The first reason is relatively straightforward, and parallels that of the reasoning for patients: certain clinicians may be more responsive to a particular type of nudge than another. For example, clinicians who are more competitive may respond better to peer-comparison nudges.

It would be interesting if, in the future, EHRs were able to test an array of different nudges for particular clinical contexts and monitor clinician activity and potential behavior change to then deliver more appropriate and personalized nudges. Knowing certain clinician characteristics (i.e. their level of expertise) can help fuel the machine learning algorithms in identifying the appropriate nudge. 

Other value additions for nudge personalization may be more discrete. Flaws in clinician decision-making extend beyond a particular treatment context. Clinicians are dealing with a variety of screening techniques, different health conditions, prescription drugs, treatment options, etc. As the body of research in this field continues to grow, it will reveal the need for many other nudges in the clinical workflow.

However, in real-world practice, implementing several nudges at a time is simply not feasible. Explicit nudges like the ones described in the examples above often require a clinician to redirect their attention. Too many such interruptions can contribute to “alert fatigue,” leading clinicians to become desensitized and ignore these nudges altogether. Tangentially, EHR alerts are predictive of burnout among the healthcare workforce, so curbing the number of alerts remains a priority.⁶

Personalization can be a potential solution to minimizing alert fatigue from nudges. By identifying which nudges are most urgent for particular clinicians, future technology can prioritize different interventions based on EHR data and clinician characteristics. For example, the active choice nudge described above may only be delivered to clinicians who consistently under-prescribe statins, or the traffic light intervention to the ~70% of neurologists that exhibit therapeutic inertia.

This approach shows promise. A 2020 study used EHR data to phenotype physician practice patterns and categorize physicians based on variables such as years of training, estimated workload, and number of patients seen.⁷ They found that an active choice EHR nudge was significantly effective in increasing influenza vaccination rates among clinicians with higher clinical workload, but had no effect among those with a lower clinical workload.

As clinician behavior changes, AI could also potentially identify when a nudge is no longer necessary to produce the desired behavior change, and progressively transition to a different nudge to target a different behavior. Even factors as simple as the time of day, or how many hours a clinician is into their shift, can pave the way for personalization. For example, as the day progresses, clinicians tend to order fewer cancer screening tests,⁸ but active choice nudging through the EHR was proven effective in increasing physician orders for colonoscopies and mammograms.⁹ Time-specific nudges can also help to decrease alert fatigue.

Nudges can encourage more guideline-concordant prescribing behavior among clinicians, but before these nudges are delivered at scale, several questions must still be answered. Future research must focus on how different nudges will interact with each other, and how this could affect clinician behavior. We also need more research to look at the long-term effects of these nudges. 

From this future research, while the intuitive next step would be for health systems to prioritize certain decision tools and EHR changes over others, I argue that personalized nudging for clinicians offers a solution with fewer trade-offs while maximizing effectiveness.

While it’s true that in an ideal world, all nudges in practice will be discrete and embedded seamlessly into the design of EHRs (i.e. default nudges), this is rather impractical for a variety of reasons. For example, some individuals may respond to other forms of nudges (i.e. active choice), and default nudges may not be feasible in certain contexts. Personalized nudging can help to identify which form of nudge is most effective for an individual clinician, and also help to identify those who are most in need of a nudge.

While this sounds rather imaginative, the technology is already being used for improving patient behavior. Why not work towards applying it to clinicians too?