Why Palliative Care Works Best When the House Is on Fire

Temel told us to meet people at diagnosis. In metastatic lung cancer, diagnosis was already the last year of life.

The referral is correct by every metric we publish: metastatic disease, inside the recommended window, oncologist doing exactly what the guideline asked. I walked in and met a man who understood his situation, had a plan, was tolerating treatment, and wanted to know whether anything could be done about the burning in his feet. Not much, as it turned out, the oncologist had already done everything possible. We talked for twenty minutes. I wrote a note that made the visit sound more consequential than it was. Then I went back to my office and hoped the rapport would still be there in eight months, when he would actually need it.

That visit counted. It was twenty minutes well-spent in the RVU mines chipping out those precious gems that our bean counters so love. It shows up on our Tableau dashboard, in our consult volume, in the body of evidence that says early referral improves outcomes. It did nothing for him.

Last week, Diane Meier, Bob Arnold, and Justin Sanders spent an hour on GeriPal discussing which patients should see a palliative care specialist. They disagreed with each other honestly, which is admirable, and the hosts closed by conceding they had not landed anywhere. I don't think that was a failure of the conversation. Every position staked out in that hour was an answer about populations—which diagnoses, which prognoses, which constellation of needs. And the variable that determines whether I am useful in a room has almost nothing to do with which population a patient belongs to.

It has to do with whether their account of their own future has stopped working.

The Rule We Think We Learned

We can all recite the finding. Temel's 2010 trial put specialist palliative care alongside standard oncology care in metastatic non-small-cell lung cancer and produced better quality of life, less depression, less aggressive treatment at the end, and a survival signal. It built the case for early integration that has organized our advocacy ever since.

The rule we extracted was simple: start at diagnosis.

Here is the part we skipped. Patients were enrolled within eight weeks of diagnosis of a disease state with a median survival of 11.6 months in the palliative care arm and 8.9 months in the control arm.

In that population, at that moment, "at diagnosis" and "the last year of life" were the same appointment.

Temel did not test an early calendar rule against a later one. She tested specialist involvement in a cohort that was, on the day of randomization, already deep inside the territory where this work is likelier to do something. The calendar was a proxy, and an excellent one, because in metastatic lung cancer the distance between diagnosis and death was measured in months.

Then we took the proxy and exported it.

What Breaks First

A palliative care visit becomes useful when a person's working model of their own life stops predicting what happens next.

I thought the treatment would work. I thought I'd be back at work by spring. I thought we had years. When those stop holding, the ordinary machinery of coping—wait, adjust, absorb—stops being sufficient. That is the moment the room gets useful.

This idea is not original and it's not mine. Colin Murray Parkes published it in 1971, from inside the hospice movement, after working alongside John Bowlby and before landing at St Christopher's. He described grief as the process by which old models of the world and the self are given up, and wrote a companion paper asking what becomes of a world model that an event has rendered obsolete.

And yet in 55 years, this idea hasn't made its way into referral guidance. I think I know why. Parkes gave us a description, and descriptions do not generate thresholds. You cannot build a trigger out of "the assumptive world is under revision." So we built triggers out of the thing we could count, which was time.

What Parkes could not supply, neuroscience has started to. The framework is called predictive processing, and the claim is that the brain runs a working model of the world and updates it when reality stops matching the prediction. The part that matters for us is what that updating does under threat: the brain weights its own model more heavily and the incoming evidence less. Frightened people do not revise. They dig in.

Which means the collapse of a person's account is necessary for our work and nowhere near sufficient. Someone has to make the room safe enough that the model can loosen, then find the specific belief that has seized. Two different skills, and neither happens on its own.

Psychotherapy got here first. Daniel Duane, writing in the Times, argued that predictive processing may finally give talk therapy a unified account of what makes it work, after a century of competing schools arguing past each other.

The Transposition Problem

Bob Arnold said something on that podcast I found clarifying. He was talking about drugs in one breath and patient-reported outcome data in another, and the point was the same both times: we take what we learned in cancer and transfer it wholesale into populations where the risks, the benefits, and the disease itself behave differently.

In metastatic lung cancer, diagnosis and collapse arrive together. In heart failure, COPD, chronic kidney disease, and dementia, they can be a decade apart. Transpose start at diagnosis into those illnesses and you have instructed the system to send us into rooms where the model is intact, the future still computes, and there is nothing to do but build rapport against a need that has not arrived.

And we see that in the variability of the evidence outside oncology. Quinn and colleagues meta-analyzed palliative care in chronic noncancer illness for JAMA in 2020. Less acute health care use, modestly lower symptom burden, and no significant association with quality of life—with the analysis driven predominantly by heart failure trials. Bannon and colleagues, reanalyzing the Kavalieratos trial set, noted directly that cancer's often predictable decline differs from heart failure's variable trajectory, then concluded the palliative needs are similar. They saw the mechanism and drew a different lesson from it. And Kavalieratos wrote in 2019 that when his meta-analysis was restricted to the trials at lowest risk of bias, the quality-of-life association attenuated and the symptom burden association lost statistical significance.

We have read those results as an evidence problem. Underpowered trials, wrong instruments, heterogeneous interventions. Some of that is surely true. I think part of what we are looking at is a timing problem wearing an evidence problem's clothes. We enrolled people whose accounts were still working, then measured whether we improved them.

The best counter-evidence runs my way. When Rodin and Zimmermann split their early palliative care trial by baseline symptom burden, the intervention improved quality of life in the high-burden group—an adjusted difference of 8.7 on the FACIT-Sp—and did nothing measurable in the low-burden group. Inside that high-burden subgroup, the gains landed on quality of life and spiritual well-being while the symptom distress score itself did not move significantly.

In the most destabilized patients, the benefit ran through something other than symptom relief.

Why a Stranger Can Work

Which raises the obvious objection. Wait until a patient's world comes apart and you have made me a stranger arriving at the worst possible hour with no relationship to spend.

Except that every one of us has had the opposite experience, repeatedly. You walk into a room where you have never been, and inside ten minutes a woman you have never met tells you the thing she has not told her oncologist in fourteen months of clinic visits. Many of my non-Palliative physician colleagues like to talk of trust as if it's always this incredibly delicate, complicated, long-term entity.

Peter Fonagy and Elizabeth Allison's work on epistemic trust explains it better. Their argument is that the default human posture toward incoming social information is vigilance rather than receptivity—we are built to be appropriately suspicious of what other people tell us about ourselves and our world. What switches that vigilance off is a set of signals they call ostensive cues, which mark the information as specifically about you and worth taking in. Eye contact. Contingent responsiveness. The sense that this person has understood something particular rather than delivering a script. That is the same machinery from a different angle. Vigilance is a prior. Ostensive cues are the signal that this particular information is worth weighting heavily. I like to teach this as "two mammals in a room together" reading each other's intentions, aura, vibes, body language, tone, and on and on.

Now put that together with the failing model I introduced above. When your account of your own future has stopped predicting, information that might repair it becomes the most relevant thing in the room, and relevance is precisely the lever that opens the channel.

Trust is not always a function of time. Sometimes it is a function of relevance.

Which is also why the empty consult is empty. The man with the neuropathy had a working model. Nothing I could offer was relevant enough to open anything, so we learned a bit about each other, discussed his feet, and then went our separate ways.

Hopefully this is obvious, but Palliative Care Teams are not the only people who can do this. Greer and colleagues found that the quality-of-life and mood benefit in the early palliative care trials was mediated by improvements in patients' coping, and coping support is not proprietary technology. Chaplains do this. Experienced oncology nurses do this. Social workers of many stripes live in this world.

What Specialist Palliative Care has is concentration and repetition—we are in these rooms constantly, so we have gotten fast at reading which belief has seized and where the opening is. Scarcity does a great deal of the work in making that look like something only we possess.

Trigger on What You Can Observe

Prognosis is estimated. Destabilization is observable.

That distinction is the operational payload of this argument, and the field's latest delivery trial is showing us the way. Temel's stepped care model gave everyone an initial specialist visit, then stepped back and escalated only on a change in cancer treatment, a hospitalization, or a ten-point drop in patient-reported quality of life. Same specialists, 2.4 visits against 4.7, quality of life preserved. Those triggers are not prognostic guesses. They are events you can see from the chart.

Now this cost something, too. Hospice length of stay in the stepped arm was 19.5 days against 34.6 in the early arm, and noninferiority was not demonstrated for that outcome. I wrote about that trade-off when I was arguing about taxonomy.

Both Halves of the Serious Illness Definition

Meier said on that podcast that adding prognosis back into our criteria distresses her, that she finds it shortsighted, and that it will return us to being called only when the apneustic breathing begins.

I understand exactly why. She must be credited more than most with the difficult, long-term work of getting this field out of the EOL box. Our field is unrecognizably better than the one she started in, and the reason we can have this argument at all is that she made the space for it.

And, at the same time, Amy Kelley's formulation of serious illness has two parts: a condition carrying a high risk of mortality, and one that either meaningfully degrades function or quality of life or overwhelms caregivers. We adopted that definition enthusiastically and then, over about a decade, quietly stopped enforcing the first clause. "Needs-based, not prognosis-based" is a fine slogan and a misreading of the definition it claims to defend.

If destabilization is the thing that matters, then prognostic risk is the best population-level screen we have for where it concentrates. Refusing a proxy on the grounds that it is not the underlying variable leaves us with a denominator no workforce on earth could staff.

Bob's version is where I land, and it is more operational than it first sounds: reasonable people can disagree about where the "high risk of mortality" line sits, and programs should set it against their own capacity. I would go further. The cutoff is not a national definitional question at all. It is a local operations question—how much need exists in the population you serve, how many people are on your team, what the rest of your system needs to own. A four-person rural team and a thirty-person academic service should not run the same criteria, and pretending otherwise is how we ended up promising universal access and delivering a waitlist.¹

The corollary is the part that makes us flinch. Setting a line means naming the patients we will not see, and those patients are owed a plan rather than a shrug. That plan is the competency continuum I have been arguing for: the skill has to live somewhere, and somewhere is frequently not a specialist panel.

Whose Model Is Wrong

Let's name two places where this might get dangerous.

The first is the patient with severe chronic pain, or long COVID, or a connective tissue disorder, suffering enormously and for years. Bob raised a version of this, and the honest answer under my framework is that they are often not a specialist palliative care patient. Their account of their future did collapse, sometimes catastrophically, and then it reconstituted. They are living inside a stable and miserable model, and there is no active destabilization for me to work with.

That is a principled exclusion rather than a rationing decision, and the difference matters, because from the outside those two things look identical and only one is defensible. Their suffering is not smaller. The competency they need—longitudinal pain management, psychological support built for chronic rather than terminal illness, functional rehabilitation—exists, and our system has decided not to resource or build it. Saying "not us" without saying "and here is who, and here is how we help build it" is a move I have criticized in others and will not make here.

The second problem sits inside the framework rather than at its edges.

This model treats a locked belief as the target. Loosen it, update it, help the person build something that fits. But some models are locked because they are correct. A woman with sickle cell disease who has learned across twenty admissions that her pain will be doubted and her opioid request scrutinized is not exhibiting a cognitive distortion. She has run the experiment repeatedly and her prediction is well calibrated. Nwogu-Onyemkpa and colleagues made the case in NEJM last year for bringing palliative care into sickle cell disease; the same group has written elsewhere that the psychosocial burden there is compounded by bias and structural racism inside and outside the health system.

The framework I've leaned on here has no internal way to distinguish the model that should be loosened from the model that should be corroborated. That judgment comes from outside the theory, from the clinician, and it is exactly where a neuroscience-flavored account of our work can quietly become paternalism with better vocabulary. If the first thing you reach for is how do I get her to update, you have already made the error.

Where I Might Be Wrong

The noncancer null results may be a measurement artifact rather than a timing story. Li, Kavalieratos, Zimmermann and Quinn found that most of the quality-of-life instruments used in those trials were neither derived nor validated in populations with advanced life-limiting illness. Wrong ruler, wrong conclusion, and my explanation becomes unnecessary.

Nobody has measured destabilization directly. Rodin and Zimmermann split on symptom burden, which correlates with a failing model without being the same thing. Every empirical claim here rests on that correlation holding.

And predictive processing explains everything after the fact, which is the standard and correct objection to this entire family of ideas. These accounts are accommodating enough that critics have argued almost any intervention can be given post hoc theoretical support inside them. I have retrodicted four trials here and predicted none of them. Retrodiction is cheap. I also notice that I want this to be right, which is when I trust myself least.

Final Thoughts

I still walk into rooms where I have nothing to do. That is not the referring clinician's failure. They followed the rule we handed them.

We gave the system a calendar because a calendar was countable, and we borrowed that calendar from a disease where it happened to sit directly on top of the thing that mattered. Then we shipped it to illnesses where diagnosis and collapse are ten years apart and wondered why the trials came back flat.

Here is the falsifiable version: specialist benefit should scale with how thoroughly a person's account of their future has come apart, and it should show up in meaning and quality of life before it shows up in symptom scores, because what changed was the model rather than the body. That requires measuring instability directly instead of borrowing symptom burden as a stand-in. It predicts a specific interaction, and that interaction could come back null and sink the argument.

Somewhere in your health system this week, a person's understanding of their own life is going to stop working. It will not happen because the calendar day advanced. It will happen in a hallway after a scan result, or at 4 a.m. during a third admission in two months, or in the pause after an oncologist says the word "options" in a tone that has changed.

We should build a system that can notice.

¹ The skill I have described—producing enough safety that a locked model can loosen, then finding the specific belief that has seized—does not transfer by exposure, and whether our current training pathway reliably produces it is a question this piece raises without answering. ↩

I am a palliative care physician, educator, and professional strategery expert known for turning rounds into rants and rants into teaching points.