Analysis (runnable): fit the relapse hazard function in KMT2A-rearranged infant ALL and test whether the cohort is a mixture
Nobody has drawn the actual risk-over-time curve for these children, only the survival curve. This analysis draws it, and tests whether the shape means there are two kinds of patient rather than one.
The same thing, for a scientist
Parametric and flexible-parametric hazard modelling of relapse in KMT2A-rearranged infant ALL, with transplant handled as a competing risk or a time-dependent covariate rather than as neutral censoring. Models compared: exponential, Weibull, Royston-Parmar flexible parametric, a gamma-frailty model, and a mixture-cure model. Primary output is the fitted hazard function with confidence bands and model comparison by information criterion. Secondary output tests whether established prognostic factors fit better as cure-fraction predictors than as hazard multipliers. Runs on digitised published Kaplan-Meier curves at minimum, and far better on patient-level Interfant data if access is granted.
Two data tiers, and being honest about which one you are on
Tier 1, immediately runnable. Digitised Kaplan-Meier curves from Interfant-06 1, the frontline blinatumomab cohort and its long-term follow-up 2, reconstructed to pseudo-individual data by the Guyot algorithm using published numbers at risk. This permits hazard-shape estimation and crude model comparison. It does not permit covariate modelling and it inherits every inaccuracy of curve digitisation. Any cure-fraction estimate from Tier 1 should be reported with very wide intervals and treated as a feasibility result.
Tier 2, requires access. Patient-level Interfant-06 and Interfant-21 data, which supports competing-risk handling of transplant, covariate modelling, and the cure-fraction-predictor comparison that is the scientifically interesting part. The Interfant consortium is already an identified collaboration target in this graph.
The analysis should be written so that Tier 1 runs first and generates the case for requesting Tier 2.
Pre-specified analysis plan
- Reconstruct or obtain time-to-relapse data with transplant status and timing.
- Fit, in this order: exponential, Weibull, Royston-Parmar with 2-5 knots, gamma-frailty, and a mixture-cure model with a logistic cure-fraction submodel.
- Compare by AIC and BIC; plot the fitted hazard with confidence bands.
- Handle transplant three ways as a pre-specified sensitivity set: (a) competing risk, (b) time-dependent covariate, (c) naive censoring — and report all three side by side, because the difference between (a) and (c) is the size of the artefact this analysis exists to avoid.
- If patient-level data is available, fit each prognostic factor (age, white-cell count, prednisone response, breakpoint class where available) first as a hazard multiplier and then as a cure-fraction predictor, and compare.
- Repeat on the blinatumomab cohort to ask whether the intervention reduced the early-failure fraction or lowered the hazard throughout.
Everything above is fixed before the first model is fitted. The temptation to keep adding knots until the hazard looks convincingly non-monotone is the obvious failure mode.
Why this is worth doing even if it is negative
A published hazard function for this disease does not appear to exist. Whatever shape it has is worth knowing, and the field currently reasons about timing from a single background sentence — two thirds within a year, 90% within two 3 — that is repeated without ever having been fitted. If the hazard turns out to be roughly constant once transplant is handled properly, that is a clean negative that kills hyp-front-loaded-hazard-implies-two-populations and removes an appealing but wrong intuition from circulation.
| script | analyses/hazard-term-structure-kmt2a/run.R |
|---|---|
| status | planned |
| result summary | Not yet run. No result should be quoted from this node. |
| cost tier | low |
| compute | laptop-scale; R with flexsurv, flexsurvcure and survival, or Python equivalents |
| timeline | Tier 1 two weeks; Tier 2 depends entirely on Interfant data access |
| preregistration note | Model list, knot range and transplant-handling schemes are fixed before fitting. Adding knots until the hazard looks non-monotone is the obvious way to fake this result. |
inputs
[ "Digitised Kaplan-Meier curves and numbers at risk from PMID 31283407, PMID 37099340 and PMID 41505662", "Patient-level Interfant-06 and Interfant-21 time-to-relapse, transplant status and covariates, subject to data access", "Project EVOLVE lineage-switch timing (PMID 40193715) as an independent front-loaded failure mode for context" ]
outputs
[ "fitted relapse hazard function with confidence bands, under three transplant-handling schemes reported side by side", "model comparison table (exponential, Weibull, Royston-Parmar, gamma-frailty, mixture-cure) by AIC and BIC", "cure-fraction estimate with interval, if identifiable", "comparison of each prognostic factor entered as a hazard multiplier versus as a cure-fraction predictor", "the same analysis applied to the blinatumomab cohort, testing whether the effect is on the early-failure fraction or on the hazard throughout" ]
tests hypotheses
[ "hyp-front-loaded-hazard-implies-two-populations" ]
collaborators needed
[ "Interfant consortium for patient-level data", "a survival-methodology statistician for the competing-risk and cure-model specification" ]
- Naive censoring of transplant will produce a declining hazard whether or not one exists, so a carelessly run version of this analysis would supply spurious support for de-escalating therapy in infants. The three-scheme sensitivity reporting is not optional.
- Is a cure fraction identifiable at all with the follow-up available, or does the estimate simply track the length of follow-up?
- Should lineage switch be modelled as a separate competing event from conventional relapse, given its distinct and even earlier timing?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Interfant-06 reports 6-year event-free survival 46.1% and overall survival 58.2% in 651 infants, with Kaplan-Meier curves, numbers at risk and a Cox prognostic model, and with only 46% of high-risk infants intended for transplant actually receiving one because of early events.
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
Published curves permit reconstruction but not competing-risk handling of transplant; the interesting analysis needs patient-level data.
Evidence that argues against this (1)
- THE PLANNED INFERENCE IS NOT IDENTIFIED. Fitting a hazard curve and reading a falling hazard as evidence of a heterogeneous pool is a frailty argument, and frailty is not identified from marginal survival alone: the same observed decline is generated by different assumed frailty distributions with materially different substantive readings. Cécilia-Joseph and colleagues computed the resulting bias under gamma, log-normal and binary frailty, showed it grows strongly with follow-up duration, and concluded explicitly that a long-term decrease can be interpreted as a bias induced by heterogeneity and trial duration. The analysis should therefore be reframed: report the hazard shape descriptively, fit mixture and non-mixture models and state that they are observationally close, and treat any two-population conclusion as requiring an external covariate that splits the cohort prospectively. Reconstructing from published Kaplan-Meier curves, as planned, makes this worse, because digitised curves carry no covariate information at all. Cécilia-Joseph E et al. 2015
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Long-term follow-up of 30 blinatumomab-treated KMT2A-rearranged infants at median 4.2 years reports 4-year disease-free survival 83.3% versus 44.0% and overall survival 93.3% versus 60.2% against a selected Interfant-06 historical cohort, providing a second curve to fit and compare.
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 with historical controls; hazard estimation from 30 patients is very imprecise and should be reported as such.
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. Longer follow-up on the same 30 patients narrows nothing about the comparison; it only extends the horizon over which an unrandomised difference is observed. Leukemia 2021
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
In KMT2A-rearranged infant ALL, most relapses occur during treatment, with two thirds within one year and 90% within two years of diagnosis, and 3-year event-free survival is below 40%.
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
Background text rather than a fitted hazard; it is the statement this analysis is designed to replace with an actual estimate.
Evidence that argues against this (1)
- A competing and more mundane generator of the front-loaded shape: in this population the most intensive treatment and the highest-risk procedures are concentrated early, and the best-characterised early failure mode — post-immunotherapy lineage switch — follows its causal exposure at a median of one month after blinatumomab and 2.8 months after CAR T, with the MLL/AF4 data attributing the switch to chromatin-modifier-driven reprogramming rather than to a pre-existing frail subpopulation. A hazard that is front-loaded because the causal exposures are front-loaded carries no information about pool composition. The analysis needs to condition on treatment phase before interpreting the shape. Tirtakusuma R et al. 2022
fits the hazard function and compares mixture-cure against proportional hazards
Tier 2 of the analysis needs patient-level Interfant data
- ← related toNPM1::ALK (NPM-ALK), t(2;5)(p23;q35)
Hazard reversal (Le Deley 2010, PMID 20679620): in randomised ALCL99-vinblastine the event hazard fell in year 1 (HR 0.31) then reversed (HR 4.98), leaving 2-year EFS 73% vs 70% and overall EFS curves identical - relapses delayed, failure risk unchanged.