Hypothesis: KMT2A-rearranged infant ALL is two populations, not one — a mixture-cure model will fit better than proportional hazards, and combination therapy only needs to be given during the first year
Nearly all relapses happen in the first year and then almost stop. That pattern usually means you are looking at two kinds of patient mixed together, not one kind whose risk fades. If so, the job is to find out at diagnosis which child is which, and to add extra drugs only during the dangerous window.
The same thing, for a scientist
Two thirds of relapses in KMT2A-rearranged infant ALL occur within one year and 90% within two [E1]. A steeply declining hazard is the standard signature of a heterogeneous pool rather than of a within-individual declining risk. We hypothesise that (i) a mixture-cure or frailty model fits Interfant relapse data significantly better than a proportional-hazards model, after handling transplant as a competing risk rather than as neutral censoring; (ii) the established prognostic factors — age, white-cell count, prednisone response, KMT2A status — predict membership of the early-failure component rather than shifting a uniform hazard; and (iii) therefore a combination partner added to menin inhibition or to chemotherapy delivers nearly all of its benefit if given during the first 6-12 months, making time-limited combination non-inferior to continuous combination and far less toxic in an infant. This is our inference from an imported reading of the hazard curve; no hazard-function fit for this population appears in the indexed literature.
The three claims have very different strengths and should be judged separately
(i) The mixture fits better. Prior 0.6. With two thirds of events in year one and 90% by year two, against a plateau thereafter, a declining hazard is close to certain; the open question is whether it survives correct handling of transplant. A high prior here, and a cheap test.
(ii) Prognostic factors classify rather than shift. Prior 0.45. Plausible and directly testable in the same model by comparing a factor entered as a cure-fraction predictor against the same factor entered as a hazard multiplier.
(iii) Time-limited combination is non-inferior. Prior 0.3, and this is the claim with clinical consequences. It requires not only that the hazard is front-loaded but that the treatment effect is concentrated in the same window, which does not follow automatically — a partner could work by preventing a late event that would otherwise be seeded early. Nothing in the hazard shape settles that, and it is the reason this hypothesis proposes an analysis rather than a protocol change.
Why this matters more in infants than anywhere else
Every month of additional therapy in a one-year-old is charged against sixty years of anthracycline cardiotoxicity, growth and endocrine disruption, and neurocognitive cost. A combination regimen that can stop at month 12 instead of running through maintenance is a materially different drug for this population than for a 69-year-old, even if the survival curves are identical. The graph already records anthracycline cardiotoxicity and growth, endocrine and neurocognitive late effects as first-class risks.
The low-hazard component, if it exists, is being treated as though it were the high-hazard one. That is not a rounding error; it is most of the cohort by 24 months.
The way this goes wrong
Informative censoring by transplant is the main threat. Patients leave the risk set because they were well enough to be transplanted, which by itself bends the observed hazard downward. Any analysis that treats transplant as neutral censoring will find a declining hazard whether or not one exists. Handling it as a competing risk, or modelling it as a time-dependent covariate, is not optional here — it is the difference between a result and an artefact, and it is written into ana-hazard-term-structure-relapse-kmt2a as a pre-specified requirement.
A second threat: protocol-adapted intensification means treatment responds to early events, so the hazard being fitted is not a natural-history hazard. This limits the mixture interpretation and must be stated in any write-up.
| statement | The relapse hazard in KMT2A-rearranged infant ALL is significantly non-proportional and declining after correct handling of transplant as a competing risk; a mixture-cure or frailty model fits better than proportional hazards; established prognostic factors act principally by predicting membership of the early-failure component; and consequently a combination partner delivers the large majority of its benefit when given only during the first 6-12 months. |
|---|---|
| origin | analogy |
| source analogy | analogy-finance-hazard-term-structure |
| prior probability | 0.45 |
| prior reasoning | Decomposed deliberately: the declining-hazard and mixture claim at about 0.6, the prognostic-classification claim at about 0.45, and the clinically consequential time-limited-combination claim at about 0.3, because concentration of hazard does not by itself imply concentration of treatment effect. The composite reflects the hypothesis as stated, which requires all three. |
| decision value | If the cohort is a mixture, the design target changes from intensifying therapy for everyone to identifying the early-failure component at diagnosis, and the duration of combination therapy in infants becomes a legitimate question rather than an assumption. |
predictions
[ "A mixture-cure model fits Interfant relapse data better than a Weibull or exponential proportional-hazards model by information criterion, with a cure fraction estimable and bounded away from zero.", "Entering age, white-cell count and prednisone response as cure-fraction predictors fits better than entering them as hazard multipliers.", "The estimated high-hazard component is enriched for RAS-pathway subclone burden and for features already linked to upfront resistance.", "Reanalysis of the frontline blinatumomab cohort shows the effect operating principally through a reduced early-failure fraction rather than a uniformly lower hazard.", "In any trial that reports it, a combination partner given for a fixed early period shows a hazard-reduction profile concentrated in the same window in which the baseline hazard is concentrated." ]
experiments
[]
analyses
[ "ana-hazard-term-structure-relapse-kmt2a", "ana-duration-adjusted-value-menin-sequencing" ]
falsifiers
[ "After handling transplant as a competing risk, the hazard is not significantly declining and proportional hazards is not rejected — the apparent front-loading was censoring.", "A mixture-cure model fits no better than proportional hazards by information criterion, or the cure fraction is not identifiable.", "Prognostic factors fit better as hazard multipliers than as cure-fraction predictors.", "A trial comparing time-limited with continuous combination shows continuous is superior, which would show treatment effect is not concentrated where the hazard is." ]
- A mixture finding would be an argument for de-escalating therapy in the low-hazard component, and de-escalation based on a statistical model rather than a measured biomarker has harmed children before. Membership must be predictable prospectively before anyone acts on it.
- Time-limited combination could simply defer relapse past the observation window rather than prevent it, and a non-inferiority study at 2 years could be reversed at 5.
- If the high-hazard component is defined by pre-existing resistance, then the children who most need a new drug are precisely the ones in whom it works least, which caps the benefit of every strategy in this graph.
- Is patient-level Interfant data accessible for a hazard-function reanalysis, or would the work have to proceed from digitised published curves with all the loss of power that entails?
- If a cure fraction is identifiable, what is its size, and does it differ between KMT2A::AFF1 and other KMT2A partners?
- Does the same mixture structure appear in adult KMT2A-rearranged AML, which would say the phenomenon is about the leukemia rather than about infancy?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
In KMT2A-rearranged infant ALL, 3-year event-free survival is below 40%, and most relapses occur during treatment, with two thirds within one year and 90% within two years of diagnosis.
Blinatumomab Added to Chemotherapy in Infant Lymphoblastic Leukemia · van der Sluis IM, de Lorenzo P, Kotecha RS, et al. · New England Journal of Medicine 2023What this evidence cannot show
A background statement in the paper rather than a fitted hazard analysis; it motivates the hypothesis but does not test it.
Evidence that argues against this (1)
- A continuum fits the prognostic data at least as well as a two-population mixture. In Interfant-06, end-of-induction MRD stratified 6-year disease-free survival into three GRADED tiers - 60.2% negative, 45.0% intermediate, 33.8% high - in 249 KMT2A-rearranged infants. A mixture-cure model posits a cured fraction and an uncured fraction; a monotone three-level dose-response in a continuous biomarker is the signature of graded residual disease burden, not of two discrete populations. The front-loaded hazard is also partly an artefact of treatment schedule, because most relapses occur ON therapy, when pressure and monitoring intensity are both highest. Stutterheim J et al. 2021
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Interfant-06 identified KMT2A rearrangement, age, white-cell count and prednisone response as prognostic factors in 651 infants using a Cox proportional-hazards framework, with 6-year EFS 46.1% overall and 20.9% in the high-risk group.
Outcome of Infants Younger Than 1 Year With Acute Lymphoblastic Leukemia Treated With the Interfant-06 Protocol: Results From an International Phase III Randomized Study · Pieters R, De Lorenzo P, Ancliffe P, et al. · Journal of Clinical Oncology 2019What this evidence cannot show
A Cox model assumes proportional hazards and therefore cannot detect or exclude the mixture structure proposed here; this is a limitation of the existing analysis, not a criticism of the study.
Evidence that argues against this (1)
- A Cox proportional-hazards framework assumes proportional hazards, which is the assumption this hypothesis is questioning. If the hazard is front-loaded because two populations are mixed, then hazard ratios from that model are averages over a period in which the ratio changes, and the prognostic factors it identifies inherit that. Statistics in Medicine 2006
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
With median follow-up 4.2 years (range 3.2-6.0), 30 blinatumomab-treated KMT2A-rearranged infants had 4-year disease-free survival of 83.3% versus 44.0% and 4-year overall survival of 93.3% versus 60.2% in a selected Interfant-06 historical cohort, providing a comparison arm in which the early-hazard component appears to have been substantially reduced.
Sustained Benefit of Blinatumomab in Infants With KMT2A-Rearranged ALL: Long-Term Outcomes, Toxicity, and Pharmacokinetics · Vieira Martins M, de Lorenzo P, Kotecha RS, et al. · Journal of Clinical Oncology 2026What this evidence cannot show
n=30 against selected historical controls. Whether the intervention shrank the high-hazard component or lowered the hazard everywhere is exactly the question the proposed model would address, and cannot be read off the published curves.
Evidence that argues against this (1)
- Thirty infants against a historical comparator, and the comparator was SELECTED from Interfant-06 rather than randomised. Every survival figure this beat draws from that study — 81.6% two-year DFS, 83.3% four-year DFS, 93.3% four-year OS — rests on the same 30 children, so quoting several of them across entries gives one small single-arm study the appearance of several. With 30 patients the confidence interval on any of those figures spans a range that overlaps the control estimate. Leukemia 2021
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Post-immunotherapy lineage switch occurs at a median of 1.5 months with 81.4% of cases inside six months, an independently front-loaded failure mode concentrated in KMT2A-rearranged disease.
Project EVOLVE: an international analysis of postimmunotherapy lineage switch, an emergent form of resistance · Project EVOLVE consortium · Blood 2025What this evidence cannot show
Timing among cases, without a denominator, so it supports the front-loading claim qualitatively rather than contributing a hazard estimate.
Evidence that argues against this (1)
- Median 1.5 months with 81.4% inside six months is measured in a cohort assembled because it switched, and early switches are more likely to be recognised as switches than late ones, which are more readily called second leukaemias. The front-loading is therefore partly an artefact of how the cases were identified. Journal of Clinical Oncology 2025
- ← related toTerm structure of risk — a steeply falling hazard curve is a statement about heterogeneity, not about time
the falling-hazard-implies-mixture inference is not identified from marginal survival data; frailty-distribution choice changes the answer and the bias grows with follow-up
- proposed by →Term structure of risk — a steeply falling hazard curve is a statement about heterogeneity, not about time
generated by reading a declining hazard as heterogeneity
- refutes →Interfant-06 — 651 infants, and the definitive answer that myeloid-style consolidation does not help
argues the Cox proportional-hazards prognostic model cannot represent the cohort structure
time-limited combination is motivated by lifetime toxicity in infants
- related to →Blinatumomab in infant KMT2A-rearranged ALL — the largest single outcome gain in thirty years
predicts the frontline blinatumomab effect acts on the early-failure fraction
fits the hazard function and compares mixture-cure against proportional hazards