The Patient of 2050

A report that redefines the physician owes an account of the patient, and this one has so far described the patient chiefly as a biological particular: a genome, an exposome, a stream of measurements, a twin. That is one kind of particularity. It is not the only one, and it may not be the kind that decides whether any of the rest of it works.

Evidence is about a class while a decision is about a member, and no hierarchy closes the gap. That gap is habitually described as a property of the data, when it is at least as much a property of the person.

Patients differ in what they will accept as a reason, in how much of the reasoning they wish to see, and in what they are trying to accomplish with a life. One asks for the number needed to treat and means it. Another asks what the physician would do for their own mother, applying a defensible standard that ranks a trusted person's integrated judgment above a summary statistic. A third dismisses the trial because it enrolled nobody like them, and is often right to. The requirement for evidence, and the interpretation of it, is itself an N-of-1 variable, and a medicine that individualizes the dose while standardizing the justification has personalized only half of the encounter.

This too is personalization, and it is better evidenced than the profession usually assumes. Work on patient heterogeneity has accumulated across medical sociology, decision science, health services research, and human-computer interaction, and it is rarely brought to bear on what the physician should become. Five findings are informative, each gathered inside the medicine we have rather than the medicine being built.

The first is more than three decades old and has not been improved upon. Degner and Sloan surveyed 436 newly diagnosed cancer patients and 482 members of the general public, asking what role each wanted in choosing cancer treatment.²³ The headline result inverted the expectation. Fifty-nine percent of the patients wanted the physician to decide on their behalf, while sixty-four percent of the public, imagining a diagnosis they did not have, believed they would want to select their own treatment. Wanting control in the abstract and wanting it while frightened are not the same appetite.

The more consequential finding is in the statistics rather than the headline: sociodemographic variables accounted for only about fifteen percent of the variance in preference. The remaining eighty-five percent lay elsewhere, in temperament, history, belief, and the accumulated idiosyncrasy that makes a person a person. Demographic category is a poor instrument for predicting what any individual wants, which is the case against stereotyping stated in the only terms that settle arguments.

The second is recent and global. The COMFORT consortium surveyed 13,806 hospital patients at 74 hospitals across 43 countries, in 26 languages, on their attitudes toward artificial intelligence in care.²⁴ The aggregate is mildly encouraging, with 57.6 percent holding a generally positive view, and the aggregate is the least interesting thing in the study. Attitudes varied by sex, by self-reported knowledge of AI, by frequency of device use and, most instructively, by health status: patients in poor health were markedly less positive than those in very good health. The people with the most at stake were the most wary, which should temper any adoption curve extrapolated from the well.

Two preferences held broadly across the sample. Asked who should make a diagnosis, 72.9 percent wanted machine and physician to reach it together with the physician taking the final decision, against 4.4 percent who wanted the machine to decide alone, and 70.2 percent wanted reasoning that could be explained to them even where explainability cost some accuracy. The sampling was nonprobability and hospital-based, so the country figures are not national estimates. The pattern across them is the durable part, and even where patients converge, what they converge on is wanting a human who remains answerable.

That observation is not an artifact of one survey. Albert Young and colleagues screened 2,590 articles and pooled the 23 that met their criteria, and reported the same result from a different method and an earlier decade: attitudes broadly favorable, reservations substantial, and a consistent preference for human supervision of whatever the machine produced.²⁵ Two independent instruments, twenty years apart, return the same answer.

The third concerns culture, and it takes the category apart. Dana Alden and colleagues put six disease scenarios of graded seriousness to quota samples of roughly 290 middle-aged urban residents in each of seven countries: Australia, China, India, Malaysia, South Korea, Thailand, and the United States.²⁶

The strongest result held in all seven, that people wanting more say for themselves wanted less for their families, and relational interdependence measured at the level of the individual predicted desired family involvement in five of the seven. The authors drew the conclusion that clinicians should set East-West stereotypes aside because variation within a culture outruns variation between cultures. That is Degner and Sloan's eighty-five percent restated at the scale of nations, and it disposes of the shortcut by which a name, a passport, or a language on the chart is taken for a decision-making profile.

The fourth measures what patients report when the question is not framed around technology at all. The Organisation for Economic Co-operation and Development collected data from 107,011 primary-care users aged forty-five and over, linked to 1,816 practices across 19 countries, on experienced quality of care, coordination, person-centeredness, and trust.²⁷ Those four belong to the relational and organizational layer of medicine, not to its technical layer. The survey shows that they can be measured comparably across health systems at scale. A profession claiming the human factor as its differentiated asset now has an instrument capable of testing the claim. That raises the cost of asserting it and lowers the cost of proving it.

The fifth cuts the other way, and this report must take it most seriously. Gale Lucas and colleagues had participants complete a health-screening interview with the same virtual interviewer. Some were told that a person was operating it, others that it was fully automated.²⁸ Those who believed they were speaking to a machine reported less fear of self-disclosure, engaged in less impression management, displayed more sadness, and disclosed more when scored by raters blind to the condition. Novelty and convenience explain none of that. What loosens the tongue is the absence of anyone who might judge.

A companion result, which the working group called artificial compassion, belongs beside it. David Chen and colleagues asked people living with cancer to rate physician-written and chatbot-drafted answers to questions about their disease, and the chatbot answers were rated the more empathetic of the two.²⁹ What studies of this kind measure is expressed empathy, which is a property of text and can be optimized like any other property of text. They do not show that a system cares, and they do not show that a patient would choose one in a room where something is at stake. Only a person can say I care and be held to it afterward.

Those results run directly against the grain of the argument this report is building. Part II will treat certain capacities as irreducibly human: presence, homological empathy, the standing of a witness who can be held to account. For some patients, on some subjects, those are the properties that make disclosure impossible.

A person who will not say aloud that they are drinking, or hearing voices, or afraid of their partner, or unable to afford the prescription, or that the pain is in a place they were raised never to name, may say it to a machine. Where culture makes a subject taboo, the machine's incapacity to judge is not a deficiency in care but the condition on which care occurs at all. The physician of 2050 will sometimes serve a patient best by not being in the room, and by designing the pathway so that what is disclosed elsewhere reaches the record, and the clinical reasoning, intact.

There is a further possibility that goes beyond the disclosure findings. The Tamagotchi effect is the formation of genuine emotional attachment to a machine, robot, or software agent, named for the 1996 handheld virtual pet whose owners grieved when it died. It is a recent instance of a much older tendency to attribute life and inner states to whatever responds to us.³⁰ Nothing about it requires sophistication. Users of ELIZA confided in a program whose mechanism they knew to be trivial. What has changed is that the artifact now converses, remembers, and is available at three in the morning, so the attachment is no longer confined to toys.

Some patients will therefore prefer an artificial interlocutor, and prefer it for reasons that are not deficits: it does not tire, it cannot be burdened, it will hear the same fear repeated a fourth time, and it never carries the expression a person cannot quite suppress. The countervailing evidence is equally firm. Across nine studies with more than six thousand participants, identical empathic responses were rated less empathic and less supportive once they were attributed to a machine, and when people were given the choice they went on choosing human empathy even while rating the machine's responses the higher of the two.³¹

The two results are not in conflict, because they concern different acts. What people want a machine for is disclosing, and what they want a person for is being cared about; disclosure is helped by the absence of a mind on the other side, and empathy takes its value from the presence of one. The design consequence for 2050 is that a system built to feel like a caring person forfeits the safety that made it worth confiding in. The patient who forms an attachment to such a system is owed a physician who knows the attachment is real, knows what it can and cannot carry, and remains the one answerable when it fails.

Those findings force a qualification on the argument of Part II. If the presence of a person is sometimes what prevents care instead of enabling it, the human advantage cannot be a general property of human presence. It is a claim about particular functions in particular contexts: who bears the consequence of a decision; who can be held to answer for it; who can sit with a patient through what follows. Those functions do not require that every step of an encounter pass through a human, and several steps go better when they do not. The chapters that follow identify what only a person can do. They do not assert that a person should do everything.

One limitation governs the whole of this evidence. Every study above was run inside the medicine we have, with the systems we have, on people who have spent their lives under episodic care. None observed a person deciding whether to accept an intervention twenty years ahead of any symptom, or arbitrating between a physician and a model that disagree, or consenting to be measured continuously for a decade. The work describes the patient of today with unusual precision. It constrains the patient of 2050 loosely at best, and it serves better as the base case from which a program of research would have to depart than as a forecast of what people will want.

Several further features of the future patient follow.

The informed patient. By 2050 most patients will arrive having already consulted a system at least as fluent as the one in the clinic, and many will be better read on their own rare disease than the generalist facing them. Some will be well informed and some confidently misinformed, and the two are not distinguishable at the door. The physician's task shifts from supplying information to adjudicating it, and the adjudication has to be done without condescension, because the patient who did the reading is frequently right.

The gradient beneath that. Health literacy and numeracy vary enormously and track education only loosely. The same statement of a ten percent absolute risk reduction lands as reassurance, as alarm, or as noise depending on who receives it. Personalizing the explanation is as much a clinical act as personalizing the molecule.

The unwilling partner. Shared decision-making is a good default and a poor universal. Some patients experience the invitation to participate as abandonment, the handing back of a burden they came specifically to hand over. Degner's fifty-nine percent were not failing at autonomy. They were exercising it.

The collective patient. In much of the world the unit of decision is not the individual. Families expect to receive the diagnosis first, to shield the patient from a prognosis, or to deliberate together. Systems built around a single authenticated user, pushing results directly to a phone, will collide with this repeatedly, and the collision will be treated as a technical defect when it is a fact about how people decide. Alden's data set the guardrail on the remedy, since the country on the chart predicts far less than the person in the chair.

The instrumented patient, and the surveilled one. Continuous monitoring is what makes N-of-1 care possible. It is also surveillance. Willingness to be measured is not uniform, and it runs lowest in populations with historical reason to distrust institutions that collect data about them. Consent to be personalized is a real variable, and declining it must remain clinically respectable rather than recorded as a compliance failure.

The patient who changes. Preferences are not stable traits. The same person wants different things acutely, chronically, and terminally, and Degner and Sloan's comparison of patients against the public is a measurement of exactly that shift. A preference captured once and honored a decade later may be a betrayal of the person now present.

The cohort fallacy. It is tempting to assume that the digitally native young will arrive at an untroubled, machine-comfortable old age in 2050. That is a stereotype about a generation rather than about a culture, no better founded than the others.

Asking is a skill with content, and three questions apply. How much of this decision do you want to make yourself, and how much do you want me to make? Who else belongs in this conversation, and should they hear it before you do or with you? And what do you want this information for, whether to plan, to decide, or to prepare someone else? Each takes seconds, and each is answerable by patients who could not describe their own decision-making style in the abstract. Each has to be asked again when the situation changes, since the answers move between the acute, the chronic, and the terminal. Recorded in the chart alongside the allergies, they are as clinically consequential as anything else kept there.

Asking well at the bedside does not settle what the profession needs to know, and four questions arise. Where a machine performs a task as well as or better than a physician, what do patients still want a physician answerable for, and on what grounds. When a model, a physician, a patient, and a horizontal care team disagree, who should advise, who should decide, who should coordinate, and who carries the accountability afterward.

Whether those preferences move with the seriousness of the situation or with the ambiguity of the evidence, both of which the existing scenarios hold roughly fixed. And how far any of it varies across cultures once variation among individuals has been taken out. Commission that work now, because preferences of this kind take longer to establish than the systems take to build, and a profession that waits will be told what patients wanted by the people who designed the interface.

None of this weakens the case for the human factor, though it adds nuance. The error to avoid is designing the physician of 2050 around a single imagined patient, informed, autonomous, individualist, data-willing, and eager for the human touch, who is largely a composite of the people who write reports like this one. The four identities set out in the chapters that follow are best understood as capacities to be deployed selectively, not as a uniform to be worn. The competence that binds them is knowing which patient wants which physician, how much of which, on this occasion and about this subject, and asking instead of inferring it from a category.

Medicine reorganizes, then, around continuous, preemptive health, a persistent, system-managed state built on N-of-1 reach, continuous monitoring, and long-horizon stewardship of each person’s trajectory.

That is the emerging reality, and Part II turns to the figure who must inhabit it.

Part Ii

The Physician Redefined