The treatments are extraordinary.
The care model is not.
A patient on a checkpoint inhibitor wakes up on Day 7 after her infusion and feels a bit off. Not enough to call. Not enough to bother the clinic. By Day 11, the diarrhea has gotten worse and the rash she ignored has spread. By Day 13, she is in the emergency room being admitted for grade 3 immune-related colitis.
This is not an unusual story. In modern immunotherapy, it is the central archetype. And it is not, primarily, a clinical knowledge problem — oncologists know exactly what to look for. It is a continuity-of-attention problem. The drugs have moved on. The model around them hasn't.
The Promise That Outran the Plumbing
In the last fifteen years, immunotherapy has done something cancer medicine waited a century for. Checkpoint inhibitors produce durable, sometimes-curative responses in tumors that used to mean a measured prognosis: metastatic melanoma, non-small cell lung cancer, head and neck cancer, renal cell carcinoma, MSI-high colorectal disease, classical Hodgkin lymphoma, and a growing list of others. CAR-T has rewritten the survival curves in relapsed-refractory leukemias and lymphomas that, a decade ago, were end-of-line diagnoses. As of early 2024, eleven checkpoint inhibitors had been approved by the FDA across more than forty distinct cancer indications, with pembrolizumab alone covering more than twenty tumor types. By 2018, an estimated 43.6% of U.S. cancer patients were already eligible for some form of checkpoint inhibitor therapy — up from less than 2% just seven years earlier — and the share has continued to expand with subsequent approvals (Haslam & Prasad, JAMA Network Open, 2019).
The clinical workflow around those drugs, however, was built for a different chemistry. Three-week cycles. A pre-cycle lab draw. A predictable nadir to watch around days 10–14. A short list of expected toxicities — nausea, hair loss, neutropenia, mucositis — that arrive on something close to a schedule. The system worked because the disease worked according to the schedule. The clinic visit was the unit of attention because the clinic visit was when the next decision needed to be made.
Immunotherapy doesn't read the schedule. And the workflow built for chemo isn't quite wrong — it's just not the workflow this drug class deserves.
Why Immune-Related Adverse Events Break the Old Model
Cytotoxic chemotherapy hits a small number of biological targets — bone marrow, gut lining, hair follicles, peripheral nerves — and largely confines its damage there. The toxicity profile is, in a sense, knowable in advance. The clinician knows when the white count will drop. The patient knows when the mouth sores will start.
Checkpoint inhibitors and other immunotherapies do something different. They release the brake on the immune system, and the immune system — once unbraked — can attack any organ it finds antigenically interesting. Skin. Gut. Lung. Liver. Pancreas. Thyroid. Pituitary. Adrenal. Heart. Joints. Kidneys. Brain. Eyes. Blood. The list is essentially every organ system in the body, and the timing window can stretch from days after the first infusion to more than a year after the last one (Postow, Sidlow & Hellmann, NEJM, 2018).
Three structural features make this drug class fundamentally harder to monitor than chemotherapy:
- Unbounded organ scope. The clinician has to be vigilant for symptoms across every system, simultaneously, in every patient.
- Wide and bimodal timing. Some toxicities arrive in the first week. Others surface months after the last infusion. There is no equivalent of the "Day 10 nadir" to anchor a watch.
- Multi-organ co-presentation. A single mild symptom is noise. A mild rash plus new diarrhea plus unusual fatigue is a multi-organ pattern that, taken together, often reads as an emerging immune toxicity. Pattern detection across organ systems is exactly what no chemo-era workflow was designed to do.
Combination regimens, which are increasingly the standard of care, magnify all three. In CheckMate 067 — the registration trial for nivolumab plus ipilimumab in advanced melanoma — 55% of patients in the combination arm experienced grade 3 or 4 treatment-related adverse events, compared with 16% on nivolumab alone and 27% on ipilimumab alone (Larkin et al., NEJM, 2015, with consistent rates across long-term follow-up). That is not an outlier number. That is the new normal in dual-agent immunotherapy.
Sources cited inline above and in the Sources section at the end of this article.
The Quiet Patient
The clinical reality above is, by itself, a hard problem. What makes it harder is what Dr. Ma identified plainly: patients on immunotherapy systematically under-report what they are feeling.
Not from indifference. From three very human reasons. They don't want to be a burden to a clinic they know is busy. They don't always know which subtle changes are clinically meaningful in this drug class — nobody told them that the rash on their chest and the loose stools after lunch belong in the same sentence. And they sometimes are indifferent in the moment, because feeling slightly off three days after an infusion is not, on its face, an emergency.
The result is a profile that almost every oncology team recognizes: the patient who turns out, in retrospect, to have had the textbook prodrome — and never said a word.
The clinical team is not at fault here. The patient is not at fault. The system around them never asked.
The patient who feels “a bit off” — not enough to call, not enough to bother the clinic — is the one who shows up days later in the ED.
— paraphrased from the central observation in Dr. Vincent Ma's interview (Misha Kaur Vilje, LinkedIn, 2026)The Cost of Silence Is Not Only the Patient's
When the prodrome is missed, the meter runs in three directions at once.
The patient pays clinically. A grade 2 colitis caught on Day 6 typically responds to oral prednisone and a hold on the next infusion. The same colitis allowed to progress to grade 3 by Day 13 is a hospitalization, IV methylprednisolone, sometimes infliximab or vedolizumab when steroids fail, and — in the worst cases — a colectomy when perforation supervenes. Pneumonitis caught early is a steroid taper. Pneumonitis caught late is an ICU admission and, sometimes, mechanical ventilation. Myocarditis caught early is rare but survivable. Myocarditis caught late is a coin flip.
The patient pays in the only currency that mattered: response. Immunotherapy works through cumulative immune engagement. A toxicity that forces an indefinite pause — or a permanent discontinuation — can interrupt the very response the patient enrolled to chase. The treatment that was working stops working not because the cancer outsmarted the drug, but because the system around the drug couldn't keep up with the drug.
The system pays in the currency that increasingly matters to it: cost. Emergency department visits, inpatient admissions, ICU stays, and second-line immunosuppression are not cheap. They are also, in aggregate, the largest avoidable line item in the economics of immunotherapy delivery. Every health system that has looked at its own data sees the same shape: a small number of late-detected immune toxicities consume a disproportionate share of the post-infusion cost curve.
This is not, primarily, a problem of clinical knowledge. Oncology has clear, well-ratified guidance from NCCN, ASCO, and SITC on how to recognize, grade, and manage every immune-related adverse event in the literature. The guidelines have done their job. What they cannot do, by themselves, is reach into the four hundred hours between infusions and ask the patient on Day 7 how she is feeling.
What the Old Model Got Right — and Why That Isn't Enough Anymore
The visit-centered, episodic model of oncology is not a failure. It is an extraordinary historical achievement: it scaled chemotherapy from an experimental hospital procedure to a community-clinic standard of care across the developed world. It worked because it matched the underlying biology. Chemo is, in important ways, an episodic drug. The visit was the right moment for the decision because the drug-induced biology obeyed a calendar.
Immunotherapy is not an episodic drug. It is a continuous one. A single infusion sets in motion an immune process that unfolds across weeks and months, in any organ, in any combination, on a timeline the clinic visit cannot anticipate. The model that worked beautifully for the first drug class is structurally mismatched to the second.
- The visit is the unit of attention.
- The patient calls when something is clearly wrong.
- The clinician watches a known, short list of expected toxicities on a known calendar.
- Most monitoring is binary — "is this an emergency, yes or no?" — with little structured way to read weak signals.
- Multi-organ patterns rarely surface, because each symptom is captured (or missed) in isolation.
- The system is reactive by design and accepts that some patients will arrive in the ED.
- The cycle — not the visit — is the unit of attention.
- The system asks. The patient answers. No new homework.
- Monitoring is structured around the actual onset windows for each organ system, not a single end-of-cycle check.
- Weak signals are read together, not in isolation. A multi-organ pattern surfaces as one story, not three forgotten messages.
- Care-team attention is routed by clinical urgency — the right human, at the right moment, with the right context.
- The system is proactive by design and treats the ED visit as the failure mode it actually is.
Five Principles for the Next Model
It is easier to say "we need a better model" than to say what one would actually look like. We've been thinking about this in conversation with oncology teams over the last year, and a coherent set of principles is starting to emerge. None of them is novel in isolation. The novelty is in insisting on all five at once.
Continuous, not episodic
Attention has to live between visits, not just at them. The cadence of the check-in should match the actual onset windows of immune-related toxicity for the regimen in question — not the calendar of the next clinic appointment. For most checkpoint inhibitors, that means a structured early-window check, a mid-cycle check, a late-onset check, and a pre-cycle review. For CAR-T it means daily attention during the cytokine-release and neurotoxicity windows. The schedule has to follow the drug, not the other way around.
Pattern-aware, not symptom-aware
A single mild symptom is noise. A mild symptom plus a faint rash plus new fatigue is a multi-organ signal that deserves a different response. The next model has to be able to read across organ systems and across days, not just within a single check-in. Pattern is the unit of clinical truth in immunotherapy. The model that only sees individual symptoms will keep missing the patients it most needs to catch.
Patient-friendly, not patient-burdening
The patient is already carrying the diagnosis, the side effects, the family stress, and the uncertainty. The next care model cannot ask them to carry the surveillance burden too. The system asks. The patient answers in their own words. The system does the rest. Anything that adds another app to learn, another protocol to remember, or another questionnaire to fill on their own initiative will simply be ignored — and the quietest patients will be the ones it ignores hardest.
Trusted clinical reasoning, not raw data or generic chat
A dashboard of vitals is not interpretation. A general-purpose chatbot is not triage. What patients and clinicians both need is a partner that can recognize this drug class, on this day of this cycle, presenting with this combination of symptoms — and frame what to do next in language each can act on. The standard for the reasoning layer is the same as for the human team: cite the guideline, explain the logic, defer when uncertain.
Routed to the right human, at the right moment, with the right context
When a signal matters, it has to reach the care team without the patient having to chase it through a phone tree, and without the team having to rebuild the story from scratch. The next model treats clinician time as the scarce resource it is — the system absorbs the volume, surfaces only what needs a human decision, and presents it with the context the human needs to act in seconds, not minutes. That is what makes proactive surveillance affordable to staff.
This Won't Be Solved by Any One Group
One of the things we appreciated most about Dr. Ma's framing — and Misha Kaur Vilje's in writing it up — is that neither of them treats this as a problem any single party will solve. It isn't a software problem; software alone cannot understand a steroid taper. It isn't a clinical-staffing problem; the workforce is already maxed. It isn't a guidelines problem; the guidelines are largely there. It is the kind of problem that gets solved when several different groups bring what they uniquely have, and agree to design together.
Bring the pattern recognition that defines what a meaningful weak signal even is.
Bring the listening infrastructure that can hold attention at scale without burning the team.
Bring the post-marketing safety datasets that should inform monitoring cadence per regimen.
Bring the aligned incentives — ED-visit reduction and on-treatment continuation are real money.
Bring the lived experience of what "a bit off" actually feels like — and what would have made them speak up.
The conversations we find most generative right now are the ones that put two of these groups in the same room and ask them to redesign one specific thing together. What does the Day-3 check-in look like for a patient on combination ipi-nivo? What does the early-warning rule for grade-2 colitis look like, expressed as something a system could actually act on? What does a payer want to see, in concrete terms, before reimbursing structured between-visit attention as a billable activity? These are not abstract questions. They have answers. We just don't yet have enough of the right rooms.
The Treatments Deserve a Care Model Worthy of Them
Twenty years ago, the most exciting sentence in oncology was "we have a new drug." Today, increasingly, the most exciting sentence is "she's been on therapy for four years and her scans are still clean." The drugs have done their part. They have expanded what cancer medicine is allowed to hope for.
It now falls to the rest of us — clinicians, technologists, pharma, payers, patient advocates, and the small companies trying to wire it together — to build the care model that lets the drugs deliver on what they have promised. Not a fancier dashboard. Not a louder reminder app. A different relationship between the patient, the clinic, and the days in between.
The patient who feels a bit off on Day 7 deserves to be asked. The clinician who would have caught it deserves to know about it before the ED does. The treatment that was working deserves the chance to keep working. And the next decade of cancer care — the decade in which immunotherapy stops being the new thing and starts being the standard thing — deserves a delivery system designed for the medicine it is actually delivering.
The drugs have been re-imagined.
Now the care model has to be.
That is the work in front of all of us. We're glad to be among the people trying to do it.
“Why immunotherapy demands a different model of cancer care”
Sources & Further Reading
- Vilje MK. “Why immunotherapy demands a different model of cancer care — interview with Dr. Vincent Ma.” LinkedIn Pulse, 2026. Read the original — the article that prompted this reflection.
- Postow MA, Sidlow R, Hellmann MD. “Immune-Related Adverse Events Associated with Immune Checkpoint Blockade.” New England Journal of Medicine, 2018; 378:158–168 — the canonical clinical review of irAE phenotypes and timing.
- Larkin J, Chiarion-Sileni V, Gonzalez R, et al. “Combined Nivolumab and Ipilimumab or Monotherapy in Untreated Melanoma.” NEJM, 2015; 373:23–34 — CheckMate 067, source of the 55% grade 3-4 treatment-related adverse-event rate in the nivolumab+ipilimumab arm.
- Wolchok JD, Chiarion-Sileni V, Gonzalez R, et al. “Long-Term Outcomes With Nivolumab Plus Ipilimumab or Nivolumab Alone Versus Ipilimumab in Patients With Advanced Melanoma” (CheckMate 067 long-term follow-up reports, NEJM 2017 and JCO 2022) — toxicity rates remained substantially consistent across extended follow-up.
- Mahmood SS, Fradley MG, Cohen JV, et al. “Myocarditis in Patients Treated with Immune Checkpoint Inhibitors.” Journal of the American College of Cardiology, 2018; 71:1755–1764.
- Salem JE, Manouchehri A, Moey M, et al. “Cardiovascular toxicities associated with immune checkpoint inhibitors: an observational, retrospective, pharmacovigilance study.” Lancet Oncology, 2018; 19:1579–1589.
- Naidoo J, Wang X, Woo KM, et al. “Pneumonitis in Patients Treated With Anti-Programmed Death-1/Programmed Death Ligand 1 Therapy.” Journal of Clinical Oncology, 2017; 35:709–717.
- Haslam A, Prasad V. “Estimation of the Percentage of US Patients With Cancer Who Are Eligible for and Respond to Checkpoint Inhibitor Immunotherapy Drugs.” JAMA Network Open, 2019; 2(5):e192535 — baseline scope estimate; eligibility has continued to expand with subsequent FDA approvals.
- National Comprehensive Cancer Network (NCCN). Clinical Practice Guidelines in Oncology: Management of Immunotherapy-Related Toxicities. Most recent version — the operational standard for irAE recognition and management.
- Schneider BJ, Naidoo J, Santomasso BD, et al. (ASCO Guideline). “Management of Immune-Related Adverse Events in Patients Treated With Immune Checkpoint Inhibitor Therapy: ASCO Guideline Update.” Journal of Clinical Oncology, 2021.
- Brahmer JR, Lacchetti C, Schneider BJ, et al. (SITC consensus). “Society for Immunotherapy of Cancer (SITC) clinical practice guideline on immune checkpoint inhibitor-related adverse events.” JITC, 2021.
Working on the same problem?
If you're an oncologist, a payer leader, a pharma medical-affairs lead, or a fellow technologist thinking seriously about between-visit immunotherapy care, we'd genuinely love to compare notes. The good answers will come from the right rooms, not the right products.
Start a conversation Our oncology focus