Proactive Identification of Palliative Care Patients: What Eight Program Leaders Have Learned
There are three truths universally acknowledged in specialty palliative care delivery: 1) inpatient palliative care produces the strongest results when consulted by the third day of admission; 2) a consistent set of variables, including diagnoses, medications, symptoms, and past utilization are strong indicators of unmet palliative care need; and 3) non-palliative care colleagues demonstrate enormous variability in recognizing such unmet need. At the intersection is the promise of automated processes to identify patients in need of specialty palliative care consultations, sometimes called ‘triggers.’
In 2018, the Center to Advance Palliative Care (CAPC) surveyed palliative care teams across the country to learn more about their experiences with standardized and automated referrals and found far more benefits than concerns. Now, as artificial intelligence (AI) and electronic health record (EHR) technology have advanced, CAPC revisits this issue via a virtual panel discussion with eight esteemed leaders:
- Nathan Moore, MD, BJC HealthCare/Washington University in St. Louis
- Ramya Prabhakar, MD, Atrius Health
- Katherine Courtright, MD, MS, University of Pennsylvania
- Jacob Strand, MD, Mayo Clinic
- Justin Woods, MD, Providence Health System
- Kat Walker, PharmD, MedStar Health
- Rachel Adams, MD, MedStar Health
- Karl Bezak, MD, HMDC, FAAHPM, University of Pittsburgh/UPMC
Proactive Patient Identification Continues to Yield Strong Benefits
Across all programs interviewed, the systematic, proactive identification process resulted in much earlier referrals to the palliative care team—and all without an increase in total consult volume (more on that below). With earlier referrals come a slew of benefits, including:
1) Better team morale
When the algorithm is based on clinical indicators, palliative care professionals felt more useful when caring for patients identified by the algorithm, and less frustrated by requests that seemed better handled by the primary team.
Dr. Strand built the Mayo Clinic algorithm based on data from patients who had previously benefited from palliative care consultations; as a result, the team sees appropriate patients and feels confident they can add value to each case. Dr. Woods (Providence) notes that the quality of palliative care consults can improve when a well-functioning alert (with built-in educational supports) expands goals-of-care conversation capacity throughout the organization, allowing palliative care specialists to direct their attention to the most complex cases.
2) Improved value
When palliative care is consulted early, patients and families have the time they need to make better-informed decisions. All programs pointed to outcomes such as reduced length-of-stay, reduced readmissions, and improved financial performance.
“Palliative care’s impact comes from improved illness understanding and prognostic awareness, but that takes time and repetition. When the referral comes late, and the patient and family are in crisis, it’s harder to have a meaningful impact.”
3) Stronger relationships within the organization
When an algorithm is designed with input from colleagues and minimizes 'false positives,' most clinicians do appreciate the nudge. This, combined with stronger financial outcomes, raises palliative care’s overall value.
Dr. Prabhakar works with her Atrius Health primary care site chiefs, who all find value in the identification process. Extensive education and proactive identification minimize the chance that someone will be overlooked in the hectic primary care environment. Proactive identification complements the regular referral process. Similarly, Dr. Courtright (University of Pennsylvania) found that the better the algorithm targets the "right" patients, the more appreciative the attendings and covering providers are of getting that alert.
And don’t forget about organizational leadership. “We need to show the impact on the metrics that leadership cares about,” notes Dr. Moore (Washington University St. Louis). “We don’t generate RVUs, so we have to show a different story.” That story becomes so much more powerful when you can document both needs AND results.
"We don’t generate RVUs, so we have to show a different story.”
Palliative Care Controls the Dials
Some palliative care leaders outside this panel have raised concerns about proactive identification, particularly the fear of being overwhelmed with consult requests. Yet the experiences of these panelists refute that, and underscore how strong, strategic implementation ensures both efficiency and sustainability.
Whether the proactive identification uses an ‘opt-in’ or ‘opt-out’ approach (see more below), the need or risk threshold for the alert, and hence the potential patient volume, can be controlled by the palliative care team.
For example, Dr. Courtright (University of Pennsylvania) worked closely with both the palliative care team and key referring clinical teams and tested the algorithm in a pilot study at her organization to determine the optimal risk threshold above which a best practice alert would prompt primary or specialty palliative care, without overwhelming the palliative care teams. Similarly, Dr. Walker (MedStar Health) found that an original alert lacking specificity was not trusted by the attendings and was more likely to be ignored; they have since improved the algorithm and settled on a relatively high threshold so that the alert is meaningful and the quality of the palliative care consult is more assured.
Washington University (Dr. Moore) set the alert to trigger goals of care conversations led by the treating team, and paired them with training, tools, and support. This initiative not only helped change organizational culture and improve the quality of palliative care referrals but also empowered the palliative care teams to triage and better manage their workloads when traditional referrals come in.
Opt-in or Opt-out?
Panelists used both approaches: ‘opt-in’ nudges the primary provider to put in a consult request, while ‘opt-out’ gives them a chance to defer a consult order before it goes to the palliative care team.
Dr. Bezak (UPMC) underscores the human-in-the-loop in the UPMC program, where the algorithm triggers not a consult but a conversation between the palliative care team and the treating team; with this nudge, the treating team often decides to conduct their own goals of care conversation—another reason why the trigger doesn't automatically increase volumes.
At Mayo Clinic (Dr. Strand), there are two opportunities for human decision-making before a consult order is placed; the algorithm’s score is first seen by a palliative care nurse, who makes the decision on whether to message the attending, and that attending can then decide whether or not to sign the prepared consult order. Atrius Health (Dr. Prabhakar) also has a palliative care nurse review the high-scoring patients, but then each PCP gets an 'opt-out’ message that says, “let us know within two weeks if you feel strongly that this patient should not be approached.”
There are good reasons for each approach:
- 'Opt-in': Both Dr. Strand (Mayo) and Dr. Courtright (University of Pennsylvania) emphasized that the primary attending has a fuller picture of the patient than any algorithm can provide and is in the best position to make an informed decision, while the alert still ensures that all patients with likely needs are considered.
- 'Opt-out': Often yields more patients receiving palliative care consults than 'opt-in' approaches, but the panelists acknowledge that this may not work in all organizational cultures.
Who Exactly Are We Looking For?
With data manipulation and AI the best they’ve ever been, programs can pinpoint patients according to the specific goals of their organization, and those goals may differ by setting or over time.
Programs can pinpoint patients according to the specific goals of their organization.
UPMC used proactive identification to find patients prior to tertiary transfer at risk of poor outcomes, while the Mayo Clinic sought patients who can benefit from palliative care services regardless of mortality risk.
The other panelists use data to find those with limited prognosis, and the existing proprietary Epic risk model is purportedly built to find patients with a poor 1-year prognosis. As Dr. Courtright (University of Pennsylvania) notes, “Many practicing clinicians still think that palliative care is for patients who are imminently dying. If the algorithm doesn’t at least consider prognosis in some way—implicitly or explicitly—those clinicians will think the alert is not correct or applicable. It is important to work with and meet referring clinicians where they are, and then progress together from there.”
"It's important to work with and meet referring clinicians where they are—and then progress together from there.”
Dr. Courtright and Dr. Woods (Providence) also see the promise of large language models (LLMs) to proactively identify patients. The advantage of making nursing notes usable for identification is that they contain indicators like nutritional or functional decline, family distress, or patients changing their minds. LLMs then let the organization prioritize patients by whatever situation they are trying to improve. “There are already AI patient monitoring systems, such as for sepsis,” Dr. Woods notes. “You can tune the AI to monitor for symptom distress or perhaps improvements in medications.”
Where Do We Go from Here?
Given the increasing capabilities of data and AI, the strong positive benefits, and the continued interest in standardizing palliative care delivery and scope, we can only expect the use of proactive patient identification processes to increase.
Drs. Courtright, Walker, and Adams are currently leading a clinical trial that identifies patients based on real-time clinical documentation of needs and compares traditional clinician referrals, an opt-in best practice alert, and an opt-out best practice alert within MedStar Health, shedding light on how outcomes and palliative care processes may vary with these three approaches to inform health systems considering implementing one. Drs. Woods and Bezak are continuing to explore LLMs and other AI technologies in palliative care, while Dr. Strand is working with Bayesian Health to disseminate their unique model to other organizations.
These leaders are working at a time when payers and accountable care organizations are also exploring and refining their own algorithms and processes to improve care for patients with serious illness.
We thank our panelists for sharing their expertise and look forward to re-exploring the status of proactive identification in a few years!
