Oncofusion

start / graph / hyp-infant-presenting-features-are-ascertainment-dominated

Hypothesisspeculative confidence 1 of 5: Our own inference or analogy. Not shown to be true.Q7-clinical-infant

Hypothesis: the spread in infant presenting-feature frequencies is mostly measurement, not biology

also called ascertainment hypothesis infant presentation · measurement dominates infant clinical picture

Reported rates for the same sign in infant leukaemia differ enormously between studies - CNS involvement from 14% to 34%, enlarged liver and spleen from 65% to 88%. The hypothesis is that most of that spread comes from how and when each study looked, not from the babies being different.

The same thing, for a scientist

Across infant ALL cohorts the reported frequency of the same presenting feature varies by a factor of two or more: CNS involvement at diagnosis 13.8% (Interfant-06) versus 33.6% (ANZCHOG) versus 25% of those tapped (TCCSG) versus 66.6% (n = 9 single hospital); hepatosplenomegaly 70%/68.3% (TCCSG) versus 87.5% (China, 144) versus 6 of 8 (single hospital); WBC >=300 x10^9/L 11.2% (CCCG) versus 29% (Interfant-06) versus 30% (MLL-10). The hypothesis is that the **dominant** source of this variance is ascertainment - who was examined, when relative to steroid exposure, with what threshold, and with what proportion of the cohort assessed at all - rather than true between-population differences in disease biology. The prediction that distinguishes it from the alternative is specific: if ascertainment dominates, then **variance in feature frequency should be larger between protocols within a region than between regions within a protocol**, and it should collapse when a single protocol's definitions are applied. If biology dominates, the reverse.

Go deeper
Why this matters rather than being a methodological aside

Three downstream uses of these frequencies are already live in this graph.

  1. Risk stratification. CNS involvement is a high-risk criterion in Japanese trials. If a third of the between-cohort spread in CNS positivity is procedural, then the same infant is high-risk in one country and standard-risk in another, and the trials are not comparing like with like.
  2. Historical controls. Every single-arm infant pilot - blinatumomab included - compares against a historical Interfant cohort. If presenting-feature ascertainment shifted between eras, the control arm's risk composition shifted with it.
  3. Diagnostic advice. A clinician told that hepatomegaly occurs in 64-88% of infants with leukaemia is being given a number whose numerator depends on an unstated centimetre threshold.

The CCCG deferred-lumbar-puncture experience is the clearest single case: the same protocol change plausibly explains both a lower CNS involvement rate at diagnosis and a lower isolated CNS relapse rate, and the data cannot distinguish treating occult disease from failing to detect it.

What would kill it

This is a falsifiable claim about variance decomposition, not a general scepticism. It dies if any of the following is observed:

  • Re-reading a single protocol's stored lumbar-puncture cytospins and flow data under one definition still leaves a two-fold CNS-positivity gap between participating regions.
  • A prospective infant cohort that fixes the timing of the first tap, the CNS classification, the organomegaly threshold in centimetres and the proportion assessed still reports feature frequencies differing by more than about 30% relative between centres.
  • A biological correlate (KMT2A partner, breakpoint, immunophenotype, or methylation subgroup) explains the between-cohort CNS or organomegaly gap after adjustment, which would locate the variance in the disease.
Facts and figures
statementMost of the between-cohort variance in reported infant ALL presenting-feature frequencies is attributable to ascertainment - timing of assessment relative to therapy, definitional thresholds, proportion of the cohort assessed, and cohort selection - rather than to true differences in disease biology between populations.
origindata-pattern
prior probability0.6
predictions
[
  "Applying a single CNS classification and a fixed tap timing to stored Interfant and ANZCHOG data narrows the 13.8% versus 33.6% gap by more than half.",
  "Organomegaly frequency recorded in centimetres below the costal margin shows far less between-cohort spread than the binary does.",
  "Within a single protocol, between-centre variance in feature frequency exceeds between-protocol variance in disease genotype composition.",
  "CNS positivity loses prognostic significance in any cohort where KMT2A status and age are both in the model, as it did in ANZCHOG."
]
analyses
[
  "ana-infant-presenting-feature-evidence-audit"
]
experiments
[]
falsifiers
[
  "A re-read of one protocol's stored lumbar-puncture material under a single definition still leaves a two-fold CNS-positivity gap between regions.",
  "A prospective infant cohort with fixed tap timing, fixed CNS classification and a centimetre organomegaly threshold still shows between-centre feature frequencies differing by more than about 30% relative.",
  "A biological correlate - KMT2A partner, breakpoint position, immunophenotype or methylation subgroup - explains the between-cohort CNS or organomegaly gap after adjustment.",
  "CNS positivity retains independent prognostic significance after adjustment for KMT2A status and age in a cohort larger than ANZCHOG's."
]
How this could bite us back
  • If presenting features are ascertainment-dominated, then risk stratification built on them misallocates therapy, and the misallocation is systematic by centre rather than random.
  • Historical-control comparisons for every single-arm infant pilot inherit the ascertainment drift, so an apparent benefit can be a change in who was counted as high-risk.
  • The hypothesis cuts against this beat's own entries: if the infant-specific frequencies replacing the pooled ones are themselves ascertainment-dominated, they are better-scoped but not more accurate.
Still open
  1. Can the Interfant trial office re-classify stored CNS data under a single definition across participating groups? That single analysis would test the main prediction.
  2. Is there any infant cohort anywhere that recorded liver and spleen size in centimetres? If one exists, the organomegaly prediction is testable immediately.
  3. How much of the apparent geographic outcome gap in Interfant-06 is risk-group misclassification arising from the same ascertainment differences?
Every source behind this entry (4)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Reported CNS involvement at diagnosis in infant ALL ranges from 13.8% in Interfant-06 to 33.6% in the ANZCHOG cohort, and the ANZCHOG cohort's CNS positivity did not survive multivariable adjustment for KMT2A status and age.

    Late effects in survivors of infant acute lymphoblastic leukaemia-a study of the Australian and New Zealand Children's Haematology/Oncology Group. · Mironova D, Saraswati CM, Downie P, Lai CY, Cook E, Carruthers V, et al. · Blood Cancer J 2023
    PMID 37752126DOI
    What this evidence cannot show

    The two cohorts differ in era, protocol and country as well as in ascertainment, so their difference is consistent with the hypothesis but does not establish it. The ANZCHOG Cox model has 61 events and its non-significant CNS hazard ratio (1.13, 0.61-2.08) could reflect power.

    Evidence that argues against this (1)
    • CNS involvement is a stated high-risk criterion in Japanese trials and is associated with worse outcome in several reports, which is hard to reconcile with a purely procedural reading; if it were mostly artefact it should not predict outcome anywhere. Stutterheim J et al. 2025
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

    CCCG-ALL-2015 deferred the first lumbar puncture to day 5 after four days of dexamethasone, reported lower CNS involvement, and reported isolated CNS relapse of 2.2% against 11.9% in Interfant-06, with an approximately 10% traumatic lumbar puncture rate.

    PMID 38634141DOI
    What this evidence cannot show

    Between-trial comparison across countries and eras with a cohort that under-enrolled high-risk infants. The letter gives no numeric CNS involvement rate at diagnosis. Read in the open-access full text at PMC11290514.

    Evidence that argues against this (1)
    • The authors interpret the same observation as dexamethasone prophase protecting against CNS relapse, which is a biological mechanism rather than a staging artefact, and their reading cannot be excluded by these data. Stutterheim J et al. 2025
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

    Reported hepatosplenomegaly frequency in infant leukaemia cohorts ranges from about 70% to 87.5%, with no cohort stating the size threshold used to call an organ enlarged.

    [Clinical analysis of 144 cases of infant leukaemia]. · Wen X, Guan X, Xian Y, Dou Y, Guo Y, Xiao J, et al. · Nan Fang Yi Ke Da Xue Xue Bao 2015
    PMID 26714909
    What this evidence cannot show

    Three cohorts across three countries and four decades, mixed lineage in one of them. The absence of a stated threshold is inferred from the abstracts; a full text might state one, and two of the three full texts were not retrieved.

    Evidence that argues against this (1)
    • Lineage differs between the cohorts (the 144-case series includes 55 AML), and organomegaly is commoner in monocytic disease, so part of the spread is plausibly real and compositional rather than procedural. Isoyama K et al. 1999
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The proportion of infants with WBC >=300 x10^9/L is 11.2% in CCCG-ALL-2015 against 29% in Interfant-06 and 30% in JPLSG MLL-10, and the CCCG authors attribute the gap to under-representation of high-risk infants in their enrolment.

    PMID 38634141DOI
    What this evidence cannot show

    Enrolment selection is itself an ascertainment mechanism, but a different one from measurement error; this evidence item supports a broader reading of ascertainment that includes who enters a cohort. The Interfant-06 and MLL-10 figures are quoted secondhand from this paper's discussion.

    Evidence that argues against this (1)
    • White cell count is the one presenting feature measured by an automated analyser with no examiner judgement, so if the hypothesis is about measurement rather than selection, this variable is the one where it should fail - and the spread here is threefold. Pieters R et al. 2019
Connected to