Of every hundred people with lupus, scleroderma or inflammatory myositis who received CAR-T cells in the largest safety cohort assembled so far, eighty developed cytokine release syndrome. That sentence has been travelling through patient groups for a month, usually without the rest of the paragraph attached: sixty-six of those eighty had a fever and nothing else, thirteen needed treatment and recovered, and exactly one patient in 108 reached grade 3. Nobody in the cohort reached grade 4.
The number is not wrong. It is simply the least useful number in the paper.
The paper, published online in The Lancet Rheumatology on September 25, 2026, does something the field has needed for three years. It takes the two blood tests that every treating centre already runs, attaches them to outcomes in 108 patients across seven hospitals on two continents, and produces a three-tier model that separates a group where fever is essentially certain from a group where it happens less than half the time. That is not a cure for CRS. It is the first evidence-based answer to a question patients ask in the first consultation and clinicians have been answering from instinct: how hard is this going to hit me.
The short version
The cohort: 108 patients with refractory autoimmune disease treated at six German centres and one in Hangzhou between March 2021 and July 2025. Median age 34; 70% women; 56% lupus, 23% systemic sclerosis, 11% inflammatory myositis.
The outcome: 80% developed CRS, but 66% of the whole cohort had grade 1 only. Grade 3 was a single patient. ICANS — the neurological complication — occurred in 3%, all mild or moderate and all reversible.
The predictors: baseline B-cell count and baseline IL-6, both independently. Each doubling of either raised the odds of CRS, by 31% and 128% respectively.
The model: patients above both cutoffs developed CRS 100% of the time; patients below both, 43%. It predicts whether, not how severe — and that distinction is the whole article.
The Cohort, and What It Is Made Of
The study is a retrospective cohort, which is the weakest kind of evidence for causation and a perfectly good kind for pattern-finding in a population nobody has yet run a randomised trial in. Seven centres contributed: six in Germany, one in Hangzhou. Everyone enrolled had failed at least two immunomodulatory therapies, which is the entry ticket to every autoimmune CAR-T programme currently running.
The disease mix matters more than it looks. Lupus accounted for 60 patients, systemic sclerosis for 25, idiopathic inflammatory myopathies for 12, and 11 had other diagnoses. These are three different diseases with three different B-cell behaviours, and lumping them into one cohort is one of the study's acknowledged weaknesses. It is also, perversely, why the result is interesting: if a single pair of biomarkers holds across lupus, scleroderma and myositis, it is probably measuring something upstream of all three rather than an artefact of one.
The treatments were not uniform either. The German patients — 72% of the cohort — received one of three CD19-directed products: zorpocabtagene autoleucel, mivocabtagene autoleucel, or zolacabtagene autoleucel. The Chinese patients, 28%, received either a dual-target CD19–BCMA construct or a co-infusion of separate CD19 and BCMA products. That difference is not a footnote; we will come back to it, because it turns out to be the most China-specific thing in the paper.
Lymphodepletion was cyclophosphamide plus fludarabine for nearly everyone. One detail in the protocol deserves attention later: before lymphodepletion, patients were held to a maximum prednisone dose of 10 mg per day, and after infusion they were switched to hydrocortisone replacement. That steroid rule is what makes the biomarker timing question so sharp.
What Cytokine Release Syndrome Actually Is
CAR-T cells are T cells retrained to kill anything carrying a particular surface marker. When they find their target in large numbers at once, they do what activated T cells do: they release cytokines, and they recruit other immune cells that release more. The result is a systemic inflammatory response that, at its mild end, looks exactly like a fever of unknown origin and, at its severe end, produces hypotension, hypoxia and organ dysfunction.
The grading system, published by the American Society for Transplantation and Cellular Therapy and used in this study, is the reason a sentence like "eighty percent got CRS" is close to meaningless on its own:
| Grade | What it looks like | What is usually done |
|---|---|---|
| 1 | Fever only — no low blood pressure, no oxygen requirement | Observation, fluids, cooling; often nothing else |
| 2 | Fever plus either low blood pressure responsive to fluids, or mild oxygen need | Tocilizumab, sometimes a short course of steroids; monitoring continues |
| 3 | Low blood pressure needing vasopressors, or oxygen needing a higher-flow device | Intensive care, tocilizumab plus steroids, close organ monitoring |
| 4 | Life-threatening: shock, or oxygen need that mechanical support cannot solve | Critical care; rare in autoimmune cohorts |
In this cohort, grade 1 accounted for 66% of all 108 patients, grade 2 for 13%, and grade 3 for one patient — 1%. The median time to onset was four days after infusion and the median duration was four days. That is a fever that arrives on a predictable schedule, lasts about as long as a bad flu, and leaves.
The neurological complication, ICANS, was rarer still: three patients, two at grade 1 and one at grade 2, all reversible. In lymphoma cohorts ICANS runs in the double digits at some centres. Its near-absence here is one of the strongest arguments that autoimmune CAR-T is a genuinely different safety proposition from cancer CAR-T, and not just the same therapy pointed at a new disease.
The Two Blood Tests
Compare the patients who developed CRS with those who did not, and three baseline measurements separate them:
| Baseline marker | Developed CRS | Did not | P |
|---|---|---|---|
| B cells (cells/µL) | 43 | 6 | 0.0002 |
| IL-6 (pg/mL) | 10 | 3 | 0.0003 |
| CRP (mg/L) | 5 | 3 | 0.042 |
| Ferritin | No significant difference | — | |
Everything else the researchers checked came back flat: age, sex, body weight, disease duration, disease type, which CAR-T product was used, how many prior lines of therapy, and the maximum number of cells infused. None of them differed between the groups.
Then multivariable regression, which asks which variables still matter once the others are accounted for, narrowed it to two. Each doubling of baseline B-cell count raised the odds of CRS by 31% (OR 1.31, 95% CI 1.06–1.67, P=0.012). Each doubling of baseline IL-6 raised them by 128% (OR 2.28, 95% CI 1.30–4.65, P=0.0001). A sensitivity analysis that removed all the Chinese patients left the associations intact, which matters because it means the finding is not an artefact of the product mix.
CRP dropped out. That is not a statistical accident, and understanding why explains the biology. CRP is not an independent inflammatory signal; it is produced by the liver in response to IL-6. Once IL-6 is in the model, CRP has nothing left to add because it is measuring the same fire from further away. The same logic explains why ferritin, a workhorse marker in lymphoma CRS, showed nothing here — in autoimmune disease the iron-loading and macrophage activation that drive ferritin up in cancer patients are largely absent.
So what are these two numbers actually measuring? B cells are the target. A high baseline B-cell count means more targets standing in front of the infused cells at the moment of first contact — more engagement, more cytokines, sooner. IL-6 is the principal effector cytokine of CRS, and a raised baseline reflects the disease's own inflammatory set point before any treatment touches it. Neither number says anything about how powerful the CAR-T product is. They describe how much fuel is already in the room.
That reframing has a consequence worth stating plainly: the peak of CAR-T expansion in the patient's blood showed no correlation with CRS at all. It is not how many soldiers you send. It is how many enemies are standing there when they arrive.
The Three-Tier Model
Using a thousand bootstrap resamples, the investigators derived cutoffs of 25.66 B cells per microlitre and 6.82 pg/mL of IL-6, then divided the cohort by how many of the two thresholds each patient crossed.
| Risk tier | Patients | Developed CRS | Rate | Of which grade 2 |
|---|---|---|---|---|
| Both above cutoff | 24 | 24 | 100% | 4 |
| One above cutoff | 43 | 35 | 81% | 6 |
| Both below cutoff | 21 | 9 | 43% | 1 |
Read the top row literally: twenty-four patients, twenty-four cases of CRS. If a patient walks in with both markers above threshold, the clinical question is no longer whether they will spike a fever but what day. Read the bottom row just as literally: twenty-one patients, nine cases, and more than half escaped it entirely.
For a treating centre this is operationally useful in a way that most biomarker papers are not. It answers a resourcing question. Which patients can be monitored on a general ward, which need a bed with continuous monitoring, which need a room near an ICU. In a field where the bottleneck is often not the cell product but the number of beds that can safely hold a patient through day four, a test that sorts 108 patients into 24, 43 and 21 is worth more than its p-value suggests.
For a patient it answers a different question, and it is worth being honest about how unsatisfying the answer is. It does not tell you that you will be fine. It tells you whether your first week is likely to be boring.
What the Model Cannot Do
Here the paper is unusually candid, and this is the part that most coverage has skipped.
The model predicts whether CRS happens. It does not predict how bad it gets. The association between risk tier and severity was flat — P=0.88, which is statistical shorthand for no relationship at all. Look again at the table above: the single grade 2 event in the low-risk tier. One patient out of twenty-one, in the group the model calls safest, needed treatment beyond observation.
What did travel with severity was something else entirely: higher body weight, and no prior use of glucocorticoids. Among the 15 patients who reached grade 2 or above, 40% had systemic sclerosis — six of the 25 scleroderma patients in the cohort. The accompanying editorial, by Catherine Shen and Christoph Ellebrecht of the University of Pennsylvania, puts the warning in one line: a low-risk profile should not be interpreted as protection against severe events. Every centre running this therapy needs the tocilizumab and the escalation pathway regardless of what the pre-treatment bloods say.
The timing problem nobody has priced in
There is a subtler issue buried in the methods, and it changes how the test should be used.
The biomarkers were drawn at the eligibility visit — that is, before immunosuppression was stopped ahead of leukapheresis. So the B-cell count recorded is not the patient's untreated B-cell count. It is the B-cell count of a patient still on whatever combination of steroids and steroid-sparing agents got them to the point of being referable. A patient on high-dose prednisone and mycophenolate may show a suppressed B-cell count because the drugs are suppressing it, not because their disease is quiet.
This is precisely why the editorial's first recommendation is to measure close to infusion rather than at eligibility. The test is cheap and repeatable; the interpretation is not portable across time points. A low reading taken in March while a patient is fully immunosuppressed tells you much less about a July infusion than the same reading taken in June.
Anyone using this model — clinician or patient — should ask a single question first: when were these numbers drawn, and what was the patient taking that week. A risk tier without a timestamp is not a risk tier.
The Belief This Study Should Retire
Among patients, and not only among patients, there is a persistent idea that the intensity of the reaction is the price of the response. Get the fever, get the remission. It is an old idea in oncology, it sounds like folk wisdom, and it is wrong.
In this cohort, remission rates were 78% in patients who developed CRS and 91% in those who did not. The difference was not statistically significant (P=0.23), which in a cohort this size means the data cannot distinguish those two numbers from each other. The plain reading is that CRS carried no signal about whether the treatment worked.
The editorial states the conclusion more elegantly than a table can: CRS is a toxic effect to prevent, not a price to pay for efficacy. For a patient weighing whether to accept a therapy with an 80% reaction rate, that sentence is the one to keep. The fever is not the therapy working. It is an adverse event that can, increasingly, be anticipated, monitored and treated — and its absence is not evidence that the cells failed.
For context on what the therapy did achieve in this cohort: 81% reached disease remission overall, and B-cell depletion lasted a median of 90 days. Those are the efficacy numbers against which the 80% should be weighed, and they are the reason this cohort exists at all.
Why Autoimmune Is Not Lymphoma
The reason CRS is milder here than oncologists expect is structural, not lucky. In a patient with relapsed lymphoma, the target compartment can be a bulky mass of malignant B cells. In lupus or scleroderma, the target is a far smaller population of autoreactive B cells distributed through blood, lymphoid tissue and, in some cases, inflamed organs. Less target, less engagement, less cytokine — the same mechanism that makes the biomarker work in the first place.
This is also why the field's early safety data looked almost too good. The first lupus CAR-T patients reported in 2021 had grade 1 fevers and little else, and the obvious worry was that the follow-up was too short or the patients too carefully selected. Four years and 108 patients later, the picture has held: a high rate of mild events, a low rate of serious ones, and a neurotoxicity profile that barely registers.
That is a genuine achievement, and it is also a reason for discipline in how the achievement is described. A 1% grade 3 rate in a cohort of 108 means, arithmetically, that the confidence interval around severe risk is wide. The authors say as much: heterogeneous population, modest sample, cutoffs not externally validated, overfitting a real possibility. The right posture is that this is the best available map, not the final territory.
Where China Sits in This
Twenty-eight percent of this cohort was treated in Hangzhou, and the design of their treatment is the detail worth carrying away.
The German patients received CD19-directed products, which is the approach that built the field. The Chinese patients received either a dual-target CD19–BCMA construct or a co-infusion of two separate products, one against each marker. The logic is that autoreactive B cells are not a single population: some carry CD19, some carry BCMA, some carry both, and a therapy that addresses only one marker leaves a reservoir behind. Hitting both is a bid for deeper, more durable immune reset.
This is where Chinese autoimmune CAR-T differs most from its Western counterpart, and the pipeline is now crowded with variations on the theme. Gracell's GC012F, a BCMA/CD19 dual-target product developed on a next-day manufacturing platform and now part of AstraZeneca, holds Chinese trial clearance across four indications including refractory lupus. Qihan's QT-019C is an allogeneic, gene-edited CD19/BCMA product with trial clearance from both the FDA and Chinese regulators for refractory lupus. Juventas has taken a CD19 product, HY001N, into registered trials for lupus nephritis and autoimmune haemolytic anaemia. And CSPC's SYS6042, cleared for a lupus trial in August 2026, is not a cell product at all in the conventional sense: it delivers CAR mRNA by lipid nanoparticle so that the patient's own T cells become CAR-T cells in the body, a format whose entire selling point is a lower rate of exactly the complication this paper is about.
Two things follow, and they point in opposite directions.
The first is that the safety question is likely to keep improving. Dual targeting engages more target cells, which on the logic of this paper should raise CRS risk rather than lower it — yet the sensitivity analysis excluding Chinese patients changed nothing, suggesting the B-cell and IL-6 relationship is robust to product architecture. Formats that avoid lymphodepletion altogether, or that express the receptor transiently, are being pursued specifically because they may decouple the therapy from the toxicity.
The second is a caution that has nothing to do with biology. No regulator anywhere has approved a CAR-T product for an autoimmune indication. Every lupus, scleroderma or myositis patient treated to date has been treated in a trial, an investigator-initiated study, or an expanded-access programme. The nine CAR-T products approved in China are approved for lymphoma, myeloma and one gastric indication; approval for cancer does not extend to lupus, and a product being studied in lupus does not make it purchasable.
For an international patient this distinction is the whole ballgame. The question is not whether China offers autoimmune CAR-T. It is whether a specific programme at a specific hospital is currently enrolling patients with your specific diagnosis, whether it accepts foreign nationals, and what the reimbursement situation is — which, for a trial, is usually that the cell product is free and the hospital care is not.
Five Questions Worth Asking
If you are considering autoimmune CAR-T and have a consultation ahead of you, these are the questions this paper makes answerable. None of them requires the clinician to have read it.
1. What are my baseline B-cell count and IL-6, and when were they drawn?
Both are routine. The tier only means something if the draw date sits close to the infusion and reflects your disease rather than your current immunosuppression. Ask for the numbers, not the conclusion.
2. If I am in the low-risk group, what changes about my monitoring?
The honest answer is: less than you would hope. The model does not predict severity, and one patient in the low tier still needed treatment. What should change is where you are observed and how fast escalation can happen — not whether anyone is watching.
3. Is tocilizumab stocked at this hospital, and who authorises it at 3 a.m.?
This is the single most practical safety question in autoimmune CAR-T, and it is asked far too rarely. The drug is standard; the protocol for releasing it without waiting for a senior review is what separates centres.
4. Does the programme lower B cells before collection, and if so how?
The paper's first proposed intervention is pre-treatment with an anti-CD20 antibody to reduce the target burden before infusion, borrowed from bridging practice in oncology. It is a hypothesis, not yet a validated step, but a centre that can explain its reasoning on this is a centre thinking about the mechanism.
5. Is this a trial, and what does participation actually cover?
Given that no autoimmune CAR-T is approved anywhere, the answer should be yes. Then ask what the sponsor pays for, what the patient pays for, and what happens to monitoring costs if CRS arrives on day four and extends the stay.
What Would Change This Conclusion
Three things, and it is worth naming them so that the next paper on this subject can be read against them rather than in isolation.
An external validation cohort, drawn from a centre that did not contribute to this one, using the same two cutoffs. Until that exists, 25.66 and 6.82 are fitted numbers, and fitted numbers tend to shrink when they travel.
A prospective study in which the biomarkers are drawn close to infusion rather than at eligibility — the design the editorial is effectively asking for — because that is the version of the test a clinic could actually act on.
And an interventional trial: does lowering the B-cell count before collection, or tightening inflammation control before lymphodepletion, actually reduce CRS without blunting the response? The authors raise both as plausible. Neither has been tested. A model that predicts a fever is a useful thing; a manoeuvre that prevents one would be a better thing, and the field is one trial away from knowing whether the second follows from the first.
A closing note on how to read numbers like these.
Eighty percent is the kind of figure that ends conversations, and the work of reading a paper properly is mostly the work of refusing to let it. What the cohort actually reports is a therapy that reliably produces a four-day fever, rarely produces anything worse, appears to work about as well whether or not the fever comes, and can now be sorted in advance into groups whose first week is predictable and groups whose is not. That is a less dramatic sentence than the headline, and a far more useful one to carry into a consultation.
Source: Chan SCW, et al. Predicting prevalence and severity of cytokine release syndrome after CAR T-cell therapy in patients with autoimmune disease: a retrospective cohort study. Lancet Rheumatol. 2026;8(11):e844–e852. Published online September 25, 2026. doi:10.1016/S2665-9913(26)00230-4. Accompanying comment: Shen C, Ellebrecht C. Lancet Rheumatol. 2026. doi:10.1016/S2665-9913(26)00279-1.