What complexity science reveals about the thing healthcare cannot stop pretending to solve — and what happens when you take that seriously in front of the people who run the system.
Uncertainty in Healthcare Through a Complexity Lens
Uncertainty in healthcare is not a gap waiting to be filled by better research. It is a structural property of the system — an emergent feature of the interaction between biological heterogeneity, incomplete knowledge, human judgement, and social context.
There is a silent contract in modern healthcare. The patient arrives with a symptom. The system responds with tests, imaging, specialist opinions, guidelines, and eventually a diagnosis. The implicit promise is: follow this process, and we will tell you what is wrong and what to do about it. The contract is deeply held on both sides. Patients need it to function. Clinicians need it to practise. The entire architecture of health technology assessment, insurance reimbursement, and clinical governance is built on the assumption that the system knows — or can come to know — what is happening inside a body, and what should be done about it.
The problem is that, for a significant proportion of patients with complex, chronic, or rare conditions, this promise cannot be kept. Diagnosis takes years. Prognosis remains probabilistic. Treatment effects vary in ways no trial has yet explained. The system responds with the best approximation it has — a label, a drug, a protocol — and the patient is left to make a life in the gap between what medicine knows and what they need to know.
This gap is uncertainty. And uncertainty, in healthcare, is both enormously consequential and almost never discussed openly.
WorkingComplexity applies complexity science to healthcare problems — drawing on complex adaptive systems theory, network medicine, dynamical systems, and implementation science to interrogate why healthcare systems behave the way they do. From this perspective, uncertainty is not a gap waiting to be filled by better research. It is a structural property of the system: an emergent feature of the interaction between biological heterogeneity, incomplete knowledge, human judgement, and social context. Understanding uncertainty through this lens means taking it seriously as information — information about the nature of the disease, the limits of the tools, and the design of the system.
Four people sit at the centre of this paper. Harold, the farmer. George, the engineer. Francis, the retired civil servant. Victor, the entrepreneur. They share a diagnosis — progressive fibrotic lung disease — and an age (in their early-to-mid seventies). Everything else about their lives, their relationships, their philosophies, and their expectations of medicine is different. What makes their stories analytically valuable is precisely that difference: they encounter the same uncertainty from radically distinct vantage points, and their reactions illuminate dimensions of the problem that aggregate data cannot.
Before meeting the personas, we establish two conceptual foundations: a typology of uncertainty, and the complexity principles that will organise the analysis.
These five types are meaningfully distinct — they require different responses, different communicative strategies, and different system designs. Collapsing them into one category ("we don't have enough data") is itself a system failure.
Irreducible randomness — the stochastic variation in disease progression that would persist even with perfect knowledge. No amount of data collection eliminates this uncertainty. It must be held, not solved.
Knowledge-dependent uncertainty — what we do not yet know and, in principle, could. The two-year average diagnostic delay in ILD is largely epistemic: tractable with better tools and earlier recognition.
Uncertainty embedded in classification systems themselves. When a patient falls into "unclassifiable ILD", the framework architecture has registered not a disease but a failure of categorisation. New data doesn't fix this; revised concepts do.
Uncertainty arising from the wider healthcare system — delays, referral patterns, funding constraints, institutional incentive structures. A patient awaiting MDT review is experiencing systemic uncertainty that has nothing to do with biology.
The lived experience of not knowing what will happen — to the body, the identity, the relationships, the future. This is the uncertainty that matters most to patients, and the one the healthcare system is least equipped to address. Harold's question is not "what is my FVC?" It is "will I see my grandchild start school?"
A complex adaptive system (CAS) is one in which many agents interact with each other and their environment, adapting their behaviour in response to feedback, and producing outcomes — emergent properties — that are qualitatively different from anything the individual agents could produce alone. ILD and PPF exemplify CAS at the biological level: fibrosis is not a thing that happens to a lung; it is what a lung does when a particular configuration of stresses, signals, and repair responses converge over time.
The WorkingComplexity principles animating this analysis: emergence (the whole cannot be predicted from the parts); feedback loops (self-reinforcing patterns shape system behaviour in counterintuitive ways); attractor states (systems settle into stable configurations — reductionist practice, passive compliance, and diagnostic labelling hierarchies are all attractors); adaptive agency (patients and clinicians learn and modify behaviour); weak signal detection (early warning signs precede major transitions but fall below current sensing thresholds); and productive uncertainty (uncertainty contains signal — the system that acknowledges it honestly gathers more information than the system that pretends it does not exist).
Harold is 74. He has built and expanded a family farm with his own hands and has managed rheumatoid arthritis since his fifties with the same pragmatism he brings to everything. When breathlessness and a persistent cough arrived in his late sixties, he attributed it to age, dust, and the RA. He is a fundamentally reactive man: he does not seek medical attention until compelled to, but when he goes, he follows instructions. His wife is well. His daughter is about to have her first child — Harold's first grandchild. His deepest value is legacy: the farm, the family, something that will outlast him.
George is 70. He entered the factory floor as an apprentice at sixteen and rose to plant manager through application and competence. He has three adult children and a generation of grandchildren whom he adores. His philosophy is direct: work hard, play hard, teach those who come after you to do the same. When his breathing began to limit his gardening, he pursued answers with the same systematic diligence he had brought to production problems. George is a man who believes that problems have solutions and that the right data, properly assembled, will reveal them.
Francis is 74. He spent forty years in government service and approached his career with the quiet pride of a man who believed that institutions matter and that serving them well was itself a form of honour. He has no children. His wife died three years ago, from a cancer that moved faster than any of them expected. He had planned a retirement of travel. His breathlessness arrived like an unwelcome bureaucratic complication. Francis processes experience through documentation, structure, and procedure. He wants to know what the pathway is.
Victor is 74 and lives expansively. He came from nothing and has spent five decades making up for it. His PPF diagnosis is, in his framing, the latest setback in a life that has had several. He is not frightened of mortality in the abstract — he has thought about it — but he resists any framing that asks him to slow down. He lavishes gifts on the people he loves and tells anyone who will listen that you only live life once. Victor wants his medical care to be a facilitating service: keep him functional, keep him in the game.
Harold suppresses the uncertainty through a plausible attribution: age, dust, RA. His framework — reactive, stoic, legacy-focused — makes early investigation unlikely. The weak signal is real but falls below his threshold for action. George begins to self-monitor with characteristic precision. His encounter with diagnostic uncertainty will be disturbing precisely because it challenges his core assumption that systematic investigation yields answers. Francis books an appointment with thoroughness that reflects his administrative instincts — he writes down his symptoms, prepares questions, trusts the system. Victor does not arrive at the GP with a list. He arrives, eventually, after his daughter insists. His uncertainty at this stage is existential from the first: not "what is this symptom?" but "what will this mean for the life I am still living?"
The symptom stage exposes the attractor states of each persona's interpretive framework. Symptoms do not speak for themselves; they are interpreted through the frameworks agents already carry. A system designed as though patients are blank receptors of medical information will systematically fail the Harolds and the Victors — two very different failure modes, same root cause.
In ILD, the GP appointment is typically where diagnostic delay is seeded. The clinical attractor state runs strongly toward COPD, heart failure, or deconditioning. Harold leaves reassured but undiagnosed — the action (going to the GP) has reduced anxiety sufficiently that the underlying uncertainty recedes. George leaves with a partial answer, a referral, and a new source of anxiety: the information he finds online includes IPF, and the prognosis is frightening. Francis finds that the appointment does not match the procedure he anticipated: "symptom → investigation → diagnosis → plan" becomes "come back if things change." Victor agrees to the referral efficiently — not because he is engaged with the process but because he wants it resolved so he can move on.
Harold absorbs ambiguous CT results without asking what "consistent with fibrosis" means in practice. The clinical encounter is calibrated to epistemic exchange (what does the scan show?) not existential support (what does this mean for the farm?). George brings a notebook, asks about sensitivity and specificity, and is disturbed to discover that the MDT will make the final assessment without him present. A decision about his body is being made by a committee he cannot access, using criteria he cannot interrogate. Francis keeps a folder. He has learned from experience — his wife's cancer — that systems can fail. Victor gets a private second CT scan, at considerable expense, and learns the same thing slightly faster. The uncertainty does not diminish. What changes is his sense that he is doing something.
Diagnosis is treated by the healthcare system as the resolution of uncertainty. In ILD, it is anything but. Harold leaves with a correct diagnosis, a prescription, and "a formless weight of uncertainty about the future that nobody has helped him map." George responds with systematic depth — he understands, in a way few patients do, that the survival statistics he reads online are population averages from trial populations unlike him in important respects. He has more information and less certainty, and the combination is destabilising. Francis receives the diagnosis alone. He goes home to an empty house and reads about IPF with careful attention, wondering what this is for and who shares it with him. Victor receives the diagnosis with a performance of equanimity, asks practical questions, and begins reframing: what restaurants have good air quality, what trips are manageable, what is worth doing now rather than later.
"The diagnostic moment is an attractor state — a stable configuration the system strains toward and patients expect. But it is not a resolution. It is the formal opening of a new uncertainty landscape."
Each adaptive strategy is a legitimate response to a complex, uncertain system. Harold narrows his world — doing what he can do well, releasing what he cannot. His wife becomes the primary care node; when she is unavailable, his care quality deteriorates noticeably. The system does not register this dependency. George engages with continued systematic thoroughness that becomes, over time, less about resolving uncertainty and more about maintaining a sense of agency in the face of its irreducibility. Francis finds, somewhat to his surprise, that the disease gives him structure. He develops a relationship with his specialist nurse that is, gradually, his closest relationship. Victor manages uncertainty through velocity — travelling carefully, hosting dinners, adjusting quietly (the restaurant table nearest the door, the afternoon rest he would never previously have admitted to needing) while maintaining the social architecture of a life lived at full volume.
A complexity-aware healthcare system would recognise these as adaptive strategies rather than pathologies. Not: how do we get them to adopt the strategy we prefer? But: what does this person's strategy need to remain viable as the disease progresses?
AI-based CT pattern recognition for ILD can identify UIP patterns with sensitivity and specificity approaching experienced radiologists. In the context of Harold's two-year diagnostic delay or the GP's single-lifetime exposure to ILD, this matters. But the AI identifies a pattern. The pattern is probabilistically associated with a diagnosis. The diagnosis is a category constructed by a human classification system built on substrates chosen for epistemic convenience rather than biological fidelity. The AI reduces epistemic uncertainty within the existing ontological framework while leaving the ontological uncertainty entirely untouched.
More problematically, AI systems trained on clinical trial populations may be systematically miscalibrated for the patients in whom diagnostic delay is most consequential — Harold's RA comorbidity, George's smoking history, Francis's age and medication profile, Victor's lifestyle comorbidities. These are precisely the patients underrepresented in training sets.
Prognostic models that stratify patients by mortality risk and exacerbation probability represent a genuine advance. Moving from "three-to-five year median survival" to an individualised probability communicates something more useful. The problem is what happens to uncertainty in the communication. George hears "72% probability of acute exacerbation within eighteen months" and processes it as an engineering specification — a fact about what will happen, rather than a fact about the best available estimate, which will be wrong for 28% of people in the relevant category and wrong for everyone in a way the confidence interval cannot capture. Francis uses the number for planning. Harold uses it for legacy scheduling. Victor rejects it as applying to other people. The healthcare system has no standard mechanism for calibrating probability communication to each of these interpretive frameworks.
The promise of AI clinical decision support is distributed intelligence made computational — aggregating signals no individual clinician could hold simultaneously. In practice, current tools recommend within existing guideline frameworks calibrated to average trial patients. An MDT that becomes an AI recommendation with human signoff is not more distributed; it is more centralised under the appearance of comprehensiveness. The specialist nurse's monthly call — Francis's most important clinical relationship — replaced by an app is not augmentation. It is the replacement of distributed, contextual, relational sensing with centralised algorithmic processing.
AI has a particular relationship with dominant narratives: it amplifies them. The system that engages honestly with uncertainty produces AI tools that engage honestly with uncertainty. The system designed to project certainty produces AI tools that project certainty, with greater sophistication and greater authority.
The most pervasive narrative: uncertainty is a function of insufficient data, and more data will progressively eliminate it. This narrative cannot account for aleatory uncertainty (irreducible by definition), ontological uncertainty (collecting more data within a flawed framework produces more data within a flawed framework), or existential uncertainty (will Harold see his grandchild start school is not a question the system collects data on). The narrative survives not because it is true but because it is useful to powerful actors — pharmaceutical companies, academic research groups, technology companies — who have strong incentives to define uncertainty as epistemic and solvable, because that is the kind their tools can address.
EBM is one of medicine's genuine intellectual achievements. The RCT has produced treatments that have meaningfully extended millions of lives. The complexity critique is not that it is wrong but that it is insufficient. An RCT in IPF enrols a carefully selected population. It produces a result valid for that population. The system then applies that result to Harold (RA comorbidity, occupational exposure), George (smoking history), Francis (cardiovascular disease and polypharmacy), and Victor (lifestyle comorbidities) — all excluded from or underrepresented in the trial. The evidence is real. The uncertainty about whether it applies to any given individual is also real. EBM, as practised, treats the former as resolving the latter. Complexity thinking insists they are independent problems.
The default model of the patient in much healthcare system design is the compliant patient: one who attends appointments, takes medications as prescribed, reports symptoms accurately, and follows advice. This narrative fails all four personas. Harold's compliance with an incorrect COPD diagnosis was rational — the system gave him something to do and he did it. George's "over-research" is the behaviour of an informed patient exercising epistemic agency. Francis's procedural management is a survival strategy for a man whose social network has been decimated. Victor's rejection of the prognostic estimate is a philosophically coherent position about how to live with irreducible aleatory uncertainty. The compliance narrative protects the system from having to adapt. It is a simplification the system cannot afford to abandon because it has not developed the capability to work with the more complex reality.
The most recent dominant narrative positions AI as the technology that will finally resolve the uncertainty that has beset healthcare. Each specific claim has evidence behind it in specific contexts. The narrative inflates them into a general claim about what uncertainty is and how it is resolved — the data sufficiency narrative at higher computational power. The uncertainty landscape in healthcare is multi-dimensional, and AI is primarily well-suited to reducing epistemic uncertainty within existing ontological frameworks. It cannot reduce aleatory uncertainty. It cannot in current form address ontological uncertainty because it is trained on the frameworks that are the source of the problem. It cannot address existential uncertainty because the signals that matter to Francis and Harold and Victor are not in the training data.
What follows is not a programme. WorkingComplexity is resistant to programmes, precisely because programmes are linear solutions to problems that are not linear. It is a set of principles for how healthcare systems might engage more honestly and productively with uncertainty.
When Harold's initial diagnosis is wrong, that is not a failure of the GP — it is a signal that the system's pattern-matching architecture is miscalibrated for atypical presentations. When George cannot get a satisfying answer about his prognosis, that is a signal that aleatory uncertainty in IPF progression is genuinely irreducible and that the communication system has not been designed to convey this honestly. Building institutional practices around honest acknowledgement of what is not known — rather than communicative practices designed to project reassuring confidence — is the first and most fundamental shift.
Harold's wife holds information about his symptom trajectory that no clinical record contains. George's notebooks hold longitudinal functional data that the MDT would benefit from. Francis's monthly specialist nurse call is his most important clinical interaction and is nowhere coded in his EHR. Victor's oxygen saturation during a dinner in his city apartment tells a different story than his clinic measurements. These are all signals. The MDT is already a form of distributed sensing — it should be understood explicitly as such, and designed to maximise signal diversity. This includes patient presence or patient data in MDT review; carer and family input as formal clinical data; patient-reported outcomes designed as dynamical data (not single-point questionnaires); and community and voluntary sector as sensing nodes.
Adaptive management designs a pathway that adjusts based on the signals the patient emits, rather than one the patient follows. For Harold, the six-month check-up schedule responds to his wife's reports rather than a fixed calendar. For George, the protocol for escalation is held loosely against the data he is generating through his own monitoring. For Francis, the monthly nurse call is the primary sensing instrument and clinical decisions respond to its signals. For Victor, the care plan is framed as a facilitation service. Adaptive management requires changing not just clinical practice but the institutional and reimbursement structures that incentivise protocol adherence over responsiveness.
Epistemic collaboration means treating the patient's interpretive framework, prior experience, and illness narrative as diagnostic data, not noise to be corrected. Personalised uncertainty communication means that Harold needs uncertainty communicated through the lens of legacy and farm management; George through systematic data and probabilities; Francis through process and procedure; Victor through facilitation and velocity. The information does not change. The communication architecture must.
AI as distributed sensing amplifier — extending sensing capacity without centralising intelligence. AI as epistemic transparency tool — presenting not just a probability but an explicit representation of the model's uncertainty, the population its training data represents, and the ways any individual patient may deviate from that distribution. AI as ontological challenge — detecting the pattern "patient presentation does not fit existing diagnostic categories" and surfacing this explicitly, rather than forcing a misfit into the nearest available label.
Longitudinal adaptive trial designs that learn from the system as it responds. Real-world evidence with dynamical capture — registry data including functional trajectories and patient-reported outcomes at sufficient frequency to support dynamical analysis. Endotype-based stratification recognising that "IPF" is a heterogeneous category. Whole-health value frameworks capturing caregiver burden, functional capacity, relational impact, and existential quality. Patient-generated evidence treated as evidence — with the methodological rigour required to make it HTA-credible.
Harold sits in a car with his daughter. He has just been told he has IPF and has been given a prescription and an appointment card. He asks what "slowing progression" means in practice. The healthcare system has given him a correct diagnosis and failed him in everything else that matters in that moment: the honest acknowledgement that his disease cannot be cured, that his future is genuinely uncertain, and that the uncertainty is not a gap that will be filled but a landscape he will have to learn to navigate.
This is not a failure of individual clinicians. It is a failure of system design — of the architecture within which clinicians work, patients are received, and evidence is generated and applied.
The practice of not knowing — holding uncertainty honestly, acting adaptively within it, and building the capacity of patients and systems to navigate it together — is not a concession to the limits of medicine. It is what the science, applied rigorously to the actual complexity of human illness, demands.
An ILD Advisory Board Session — Transcript and Synthesis
The paper Living with Not Knowing was circulated to six board members 72 hours before this session. Members were asked to read in full and come prepared with a written initial reaction. What follows is the transcript of that session, lightly edited for clarity. Positions recorded here supersede positions held at previous meetings where explicitly revised.
| # | Name | Role | Organisation |
|---|---|---|---|
| 1 | Dr. Sandra Kowalski | System Chief Pharmacy Officer; Chair, P&T Committee | MidWest Health Alliance — Chicago, IL |
| 2 | Marcus Webb | President & CEO, Academic-affiliated nonprofit system | Elevance Regional Health — Raleigh–Richmond corridor |
| 3 | Janet Flores | SVP, Employer & Commercial Market Solutions | Centurion Health Plans — 6-state regional MCO |
| 4 | Prof. David Osei | Professor of Medicine; Director, ILD Program; PI | UCSF — Division of Pulmonary, Critical Care & Sleep |
| 5 | Dr. Priya Nair | Pulmonologist — ILD Clinic + General Respiratory | Desert Southwest Pulmonary Associates — Phoenix, AZ |
| 6 | Carmen Reyes | Executive Director — IPF patient (dx. 2019) | Breathe Forward Alliance — Austin, TX (12,000 members) |
Facilitator: I want to hear from each of you before we open to discussion. What is your honest, unfiltered first reaction? What lands, what doesn't, and what did it change — if anything — for you?
The five-part taxonomy of uncertainty is the most useful framework I've encountered in this space, and I say that having read a lot of complexity literature that never quite gets off the philosophy shelf. Aleatory, epistemic, ontological, systemic, existential — those are five distinct operational problems. My P&T committee has been treating all of them as one problem for twenty years, and that conflation is exactly why our evidence requirements don't actually produce better access decisions.
Where I push back: the WorkingComplexity way forward gives me six principles and no decision architecture. "Treat uncertainty as signal, not failure" — fine. Tell me what the signal means when I'm reviewing a formulary submission for a drug with a mechanism-of-action story, no real-world data, and a manufacturer who says the RCT didn't capture the right outcomes because the RCT framework is flawed. That is the room I'm going to be in next quarter. The paper tells me why I'm in a difficult room. I need more help with what to do when the door closes.
One thing that genuinely surprised me: the characterisation of Victor as the patient who drives the most expensive care is a claim I've always made through a utilisation lens. Reading Victor as an adaptive agent with a philosophically coherent response to aleatory uncertainty rather than a behaviour management problem — that reframe is going to cost me some assumptions.
I want to start where Sandra ended, because Victor is the persona I found most challenging. In a system CEO context, Victor is the patient my care coordinators flag as "difficult to engage." He refused structured disease education. He didn't want the pulmonary rehab referral. He found out his prognosis and went to a restaurant instead of scheduling the follow-up. We would categorise him as low-adherence and high-risk, and we would build a utilisation management programme around him.
The paper is asking whether that's wrong. Not just incomplete — actually wrong. That question doesn't have a quick answer, but it has significant implications for how I design care management programmes in my COE.
What landed most directly was the observation about Harold's wife as an unrecognised care node. My workforce plan accounts for the ILD coordinator we don't currently have, the subspecialist fellow we want to recruit, the advanced practice provider who will extend clinic capacity. It accounts for zero informal caregivers, even though — if this paper is correct — those caregivers hold clinical information my team will never collect. That is a gap in my system design, not a gap in the patient's behaviour.
The aleatory versus epistemic distinction is the single most operationally useful thing I've read this year — and I want to explain why, because it may not be obvious. My actuarial team prices uncertainty as a single variable. We build a rate for ILD patients based on historical claims trajectory and we call that our best estimate of future cost. What we're actually pricing is a mixture of aleatory uncertainty — the genuine unpredictability of disease progression, which no data will eliminate — and epistemic uncertainty — what we don't know because the data doesn't exist yet, which could in principle be reduced. We treat them identically in the model, which means we're systematically mispricing risk and, probably, creating access barriers for exactly the patients who would benefit most from better data.
What troubles me is the compliance narrative section. Not because it's wrong — it's broadly right — but because my employer clients have a twelve-month renewal cycle and a P&L they have to close. The right framing and the feasible framing are not the same framing, and the gap between them is where good ideas go to die.
The ontological uncertainty section is scientifically precise in a way that I haven't seen outside the technical framework architecture literature. The observation that "unclassifiable ILD" is not a disease category but a framework failure is correct and is under-discussed in the clinical literature. We don't talk about it honestly because our guidelines depend on categories that we know are leaky, and admitting the leak requires admitting that the evidence we've built on top of those categories is also, to some degree, leaky.
The ILA situation is the live demonstration of the paper's argument. We have hundreds of thousands of CT findings that don't fit existing classification architectures. The paper is asking whether we should be revising the framework rather than generating more data within it. That is the right question.
Where the paper is insufficiently hard on itself: the critique of EBM is too broad to be actionable. RCTs in IPF have produced effective treatments. The problem isn't the RCT; it's the endpoint selection and the population specification. I would also note that the claim that AI primarily reduces epistemic uncertainty within existing frameworks will be outdated within five years. Multi-omic AI approaches are beginning to generate endotype hypotheses from data patterns no existing clinical framework anticipated — AI may, with appropriate design, become a framework-challenging tool rather than merely a framework-applying one.
I want to say something that's going to sound simple but isn't: all four of those patient personas were in my clinic last week.
Harold was there. He came in because his wife made him come. He was reluctant, followed instructions, didn't ask what anything meant. I spent twelve minutes with him. I don't know what he went home and worried about because he didn't tell me and I didn't have time to ask.
George was there. He brought printed research. He had highlighted sections. He wanted to know why his FVC had changed by twelve millilitres and whether that was meaningful. I told him it was within measurement variability and I could see he didn't believe me, and I didn't have time to explain why I wasn't certain either.
Francis was there — alone. Meticulous, prepared, alone. He shook my hand at the end and said "thank you very much" in a way that made me want to ask if he was alright. I didn't. I had to see the next patient. Victor wasn't there because Victor goes private.
The paper describes my clinical reality with more precision than anything I've read from a health system design perspective. And then it gives me six principles and no 12-minute-visit translation. What I need is a concrete, evidence-based script for talking to patients about uncertainty in a way that names it honestly without overwhelming people who are already overwhelmed.
I've been in rooms where someone presents a framework paper and everyone finds their own work in it and nods and nothing changes. I don't want this to be that room. So I'm going to say a few things that I hope create productive discomfort.
The existential uncertainty section is the truest part of this paper. And it is the part that every person in this room, including me, is most likely to treat as a nice paragraph that supports the clinical argument without generating any concrete commitment to doing something about it. I'm naming that in advance so we don't let it happen.
What the paper does not say — and I want to add it here — is that the five types of uncertainty are not equally distributed. Harold has a wife. Francis does not. Victor can pay for a private CT scan. My members in rural Texas cannot. George has the literacy to interrogate his prognosis. Many of my members who are Black, Latino, or working in the agricultural and construction industries that generate occupational ILD exposure — those members are facing the same uncertainty with fewer resources and deeper consequences of getting it wrong.
I would like the board to sit with the following question before we move to round two: if you had to choose one of the four personas as the patient your current system is actually designed to serve well, which one would it be? And what does your answer tell you about who your system is not designed for?
Facilitator: Carmen has asked a direct question. Let's take it. Who is your system designed to serve well — and who is it not?
My system — the academic ILD programme — is designed to serve George. He brings printed research, asks about measurement variability, wants to understand the probability distributions behind his prognosis. I can have a long, sophisticated conversation with George. I find it satisfying.
The person my system is hardest on is Harold. He arrives already deferential to the institution. He leaves with a correct diagnosis and a prescription and an appointment card, and he goes home not knowing what any of it means for the farm he is still trying to run. I have a seventeen-point checklist for my ILD clinic intake. None of the seventeen points ask about the farm.
Carmen's redistribution point cuts me. The ILA natural history study I'm designing is primarily recruiting from centres like mine. The patients most likely to develop ILA-to-IPF progression in occupational exposure contexts are least likely to be in my study. They're not just missing from my data. They're missing from my concept of what the disease is.
My system is designed for Janet's patient. The employer group member with commercial insurance, a specialty pharmacy benefit that works, and a prior auth pathway that — with effort — can be navigated. That patient is George, maybe Francis if Francis has good supplemental coverage.
My system is not designed for Harold. His occupational exposure is relevant to his fibrosis and almost certainly never made it into his medical record in a way that my prior auth algorithm can account for. My P&T criteria are based on clinical trial populations. Harold is already outside the edge of the evidence base his treatment is justified by. I'm approving coverage for something that the trial didn't enrol his phenotype in, and I'm doing it because the alternative is denying coverage to a man who is dying.
On productive uncertainty operationally: if I accept that uncertainty is signal, then what signals is my P&T process currently discarding? For aleatory uncertainty, demanding data is correct. For ontological uncertainty — submissions for drugs targeting populations the current framework can't adequately categorise — the discount may be wrong. I'm not sure yet. But it's a question I'll take back to my committee.
My system is designed for Francis. The civil servant with good procedural instincts, who researches his appointments, follows the pathway, shows up prepared. The system was designed by people who navigate systems — with folder management and appointment structures and information hierarchies that make sense to someone with administrative literacy.
Carmen's point about informal care networks is landing differently for me now that Priya has said all four patients were in her clinic last week. I've been thinking about Harold's wife as a workforce design insight. But Harold's wife is not present in my system at all. She doesn't appear in the EHR. The information she holds doesn't have a data field. If she died — like Francis's wife died — Harold's care quality would deteriorate and my system would have no sensor to detect that deterioration until it showed up as a hospitalisation.
I designed my system for no one in particular, which is almost worse than designing it for George. I designed it for the actuarial population mean, which doesn't exist.
Here is the concrete thing I want to do as a result of this conversation: I want to test whether informal care network status is a meaningful predictor of claims trajectory. Not caregiver presence as a demographic variable — too crude — but something closer to what the paper calls "sensing capacity." Is there a signal in existing data that distinguishes Harold-with-wife from a hypothetical Harold-without-wife? If there is, I need to price that risk differently. And if I price it differently, I need to build a benefit that reduces it — which means I'm suddenly in the business of supporting caregivers, not just insuring patients.
On Priya's 12-minute script request: I would fund that development. A structured, evidence-tested, brief uncertainty communication tool at diagnosis would reduce the anxiety-driven over-utilisation I currently see in newly diagnosed patients — the second CT scans, the second opinions, the ER visits that aren't acute exacerbations but are acute anxiety. That has a return in my model within eighteen months of adoption.
I want to respond to what David said about his seventeen-point checklist. I have a checklist too. And I know exactly what it doesn't ask, because I wrote it during a three-week period when I was also managing sixty active IPF patients and doing my own prior auth submissions and I did not have the bandwidth to think about what the patient was going to go home and worry about. I wrote a checklist that captures the clinical information I need to manage the disease. I did not write a checklist that captures the information the patient needs to manage their life with the disease. Those are different documents.
The paper distinguishes between the system's epistemic needs and the patient's existential needs. I experience that distinction as a daily practical problem: my visit is reimbursed for what I document clinically, not for what I manage emotionally or relationally. If I spent fifteen minutes talking to Harold about the farm and what prognosis means for his legacy planning, I cannot bill for that. If I ordered a CT scan I may not need, I can.
Sandra raised commissioning a 12-minute uncertainty script. Janet said she'd fund it. I will develop it, on two conditions: it must be tested in community practice, not academic centres, and it must be co-developed with patients. Not reviewed at the end — co-developed from the first draft. Carmen, I want Breathe Forward's members in the room when we write it.
I'm going to answer my own question, and then I'm going to say something important about how we leave this room.
The system designed to serve my members well doesn't exist yet. It has fragments. Priya's clinic is a fragment. Marcus's COE investment is a fragment. Sandra's willingness to interrogate her P&T criteria is a fragment. Janet's actuarial rethinking is a fragment. David's ILA natural history work is a fragment. What's missing is the architecture that holds the fragments together.
The section on productive uncertainty describes what patients and caregivers in peer support groups do every day without institutional support. The twelve thousand people in Breathe Forward are practising productive uncertainty as a survival strategy. They are sharing information about their disease trajectories, comparing notes on prior auth experiences, developing collective intelligence about what works and what doesn't. That is distributed sensing. That is emergent knowledge. It exists entirely outside the formal care system and is not captured, credited, or compensated by any part of the system represented in this room.
What I want, as a concrete output from today: I want the board to acknowledge that Breathe Forward's data — four years of patient-reported outcomes, adherence patterns, and experiential evidence from twelve thousand members — is a clinical asset. Not a nice-to-have patient voice add-on. A clinical asset that should have a formal role in how this board makes decisions. Contractual representation with specified data contribution and specified influence on outcomes.
Facilitator: You've heard each other. What has changed — specifically, concretely, by person? What are you taking back to your organisation that you would not have taken back before this session?
First: I am adding uncertainty characterisation as a formal criterion in my P&T review process, starting Q3 2026. Every formulary submission will be required to identify, for the disease in question, which of the five uncertainty types is most operationally significant — and to specify what the submission does and does not address. Second: I am revising my working assumption about AI-based evidence packages — now requiring AI submissions to specify what uncertainty type the AI addresses, what it does not, and what care relationship it supplements or replaces. Third: I am calling Carmen next week about Breathe Forward's adherence data. I will find out whether it can be formalised in a way my P&T accepts.
Immediate: I am commissioning a mapping of informal care network density across my current ILD clinic population — every active patient. I want to know who has a Harold's wife and who doesn't. Once I have it, I will use it to stratify care coordination investment. Strategic: I am beginning a conversation with community affiliates about what sharing coordinator capacity would actually require contractually and operationally. My hub-and-spoke model needs to be a sensing architecture — receiving signals from the spokes — not just a referral network.
In Q3 2026, I am asking my actuarial team to develop a prototype model that separates aleatory from epistemic uncertainty for the ILD/PPF population. I am following through on convening the joint outcomes working group — with Carmen's condition (contractual representation, data contribution rights) and Priya's condition (tested for operability in community practice). And I am committing the funding for the 12-minute uncertainty communication script — Carmen and Priya lead, with Breathe Forward patient co-design, tested in three different community settings before any formulary policy is attached to it.
I am taking the EBM/evidence architecture distinction I articulated in my initial reaction to the guidelines committee I sit on. I am incorporating the five-dimensional uncertainty taxonomy as a formal measurement framework in the ILA natural history study design — adding structured measurement of epistemic, systemic, and existential uncertainty as secondary endpoints. I am formalising the community advisory panel commitment: Carmen receives a letter of intent by end of this week, with voting rights on enrolment criteria and patient-relevant outcomes specification. And I am revising my seventeen-point intake checklist to add a question about what matters most to the patient about the next twelve months that the disease might affect. Those answers belong in the clinical record.
I am developing the 12-minute uncertainty script — Janet is funding it, Carmen is co-designing it, and I will not let this dissolve into a committee. I want the script to give patients language for not-knowing that is both honest and survivable. I am following through on the ILA registry commitment — structured documentation of ILA patients even before treatment indication, because the data I collect about Harold's trajectory is not administrative overhead, it is a signal. And I am stating publicly: if Marcus's hub-and-spoke model includes genuine resource-sharing — coordinators, prior auth support, monitoring tools — I will affiliate. If it moves complexity to the community without moving the resources, I will call it out.
1: I am formally proposing Breathe Forward's longitudinal patient data as a formal contribution to Janet's working group — with co-authorship on publications, contractual governance representation, and veto rights over framings that remove patient context to make it actuarially legible. 2: I am commissioning a listening exercise with Breathe Forward members specifically on the five types of uncertainty — which are most debilitating, which they have strategies for, what those strategies are. 3: I am writing to every manufacturer with an ILD pipeline asset requesting a meeting on trial design equity. 4: I am using the existential uncertainty section of this paper as the redesign brief for Breathe Forward's peer support programme — reframed from "coping with IPF" to "navigating uncertainty together." 5: I want David Osei's letter of intent for the community advisory panel by Friday. That is not negotiable, and David knows it isn't, and that is why I said it in front of everyone.
The vocabulary shifted. Before this session, the board had no shared language for different types of uncertainty. After it, five distinct types are available as operational concepts. Problems previously discussed as "not enough data" can now be identified as distinct problems requiring distinct responses.
The patient personas did something the clinical literature doesn't. Harold, George, Francis, and Victor made visible what aggregate outcomes data averages away: that uncertainty is not experienced uniformly, that adaptive strategies are diverse and coherent, and that the system's current response to non-standard adaptation is misclassification as non-compliance. Every board member was able to name which patient they are currently designed to serve, and every board member identified gaps.
The informal care network as clinical asset. Marcus Webb's observation about Harold's wife — that she holds clinical information the formal system cannot access — generated more immediate consensus than any position paper in recent memory. The infrastructure consequence remains unsolved. But the problem has been named.
Carmen Reyes's challenge produced accountability. The question "who is your system designed for?" was uncomfortable and productive. Every answer contained an honest acknowledgement of a gap. Named problems have owners. That is where system change begins.
| Commitment | Owner | Timeline |
|---|---|---|
| P&T uncertainty characterisation criterion | Kowalski | Q3 2026 |
| Contact Breathe Forward on adherence data | Kowalski | This week |
| Informal care network mapping (ILD population) | Webb | Q3 2026 |
| Hub-and-spoke sensing architecture specification | Webb | Q4 2026 |
| Aleatory/epistemic actuarial prototype model | Flores | Q3 2026 |
| Joint outcomes working group (with patient conditions) | Flores + Reyes + Nair | 60 days |
| 12-minute uncertainty script (funded, co-designed, community-tested) | Nair / Reyes / Flores | 12 months to pilot |
| Five-uncertainty-type measurement in ILA study protocol | Osei | Before study launch |
| ILA community advisory panel — letter of intent | Osei | Friday 29 May 2026 |
| Revised intake checklist + patient question | Osei + Nair | Q3 2026 |
| ILA patient registry documentation (structured) | Nair | Q3 2026 |
| Breathe Forward data package — formal contribution proposal | Reyes | Before working group |
| Member listening exercise on five uncertainty types | Reyes | 90 days |
| Manufacturer letters on trial design equity | Reyes | 30 days |
| Peer support programme redesign (existential uncertainty frame) | Reyes | Programme cycle 2027 |
"Who is responsible for the architecture that holds the fragments together? And who in this room has the standing to propose it?"