Biomarker Cut Points
Turning a continuous biomarker into categories by choosing a threshold — a step that is statistically dangerous in a specific, well-understood, and very common way.
What it is
Continuous markers (Ki-67 percentage, tumour size, a model’s predicted probability) get dichotomised constantly, because clinical decisions are binary. Three things go wrong.
Information loss. Dichotomising discards within-group variation and reduces statistical power. A case just above the threshold is treated identically to one far above it. As a default, keeping the variable continuous and modelling non-linearity with splines is statistically superior — the categorisation should be justified by decision-making needs, not convenience.
Optimal-cut-point overfitting. This is the serious one. Searching all possible thresholds and picking the one with the smallest p-value guarantees an optimistic result: you have run dozens of implicit tests and reported the best. The resulting cut-point is unstable — it frequently fails to replicate — and the effect size at that cut-point is biased upward. This “optimal cutpoint” problem is well documented in the biomarker literature and remains widespread.
Threshold-dependent conclusions. Different published cut-points for the same marker are a recurring source of irreproducibility across studies.
The defensible options: use a pre-specified threshold from prior literature or a guideline; or derive one but validate it on split-sample or cross-validated data and report the correction; or keep the variable continuous and report the functional form. In all cases the derivation method belongs in the methods section.
The leading worked example is Ki-67 — the marker named in the opening line of this page. Different published Ki-67 cut-points are a standing source of irreproducibility, which is exactly the case for reaching for a pre-specified guideline threshold rather than an internally derived one. The UK Ki-67 recommendation (sources/papers/shaaban-2026-uk-ki67-recommendations.md) is that guideline for the UK setting; whether it endorses a specific numeric cut-point, and how it derived one, is [unverified] from the abstract, and that is precisely the detail this page says must be stated for a threshold to be defensible. See Ki-67 Proliferation Index for the marker itself.
A caveat to this page’s own advice, added 2026-08-12. “Use a pre-specified threshold from a guideline” is listed above as a defensible option, and it is — but only if the guideline derived its threshold rather than agreed it. Aust 2010 — Serrated Polyp Diagnostic Criteria (German consensus) is a national consensus that supplies a diagnostic threshold with no derivation at all: two of four architectural features, present in at least two crypts, to call a sessile serrated adenoma. No sensitivity, no specificity, no comparison against one-of-four or three-of-four, and the paper says outright that the rule is “pragmatic” and that no quantitative criteria exist.
It is not a cut on a continuous variable, and the difference is worth stating rather than blurring: this is a counting rule over qualitative features, so information loss and optimal-cutpoint overfitting in the strict sense do not apply — nobody searched a threshold space, which is why it cannot be overfitted. It is arguably the opposite failure. A searched cut-point is optimistic because it was tuned to data; this one was never exposed to data at all, so its operating characteristics are simply unknown rather than inflated.
What does carry over is the third item above, threshold-dependent conclusions, and the instability near the boundary that Mitotic Count describes: a lesion showing exactly one qualifying feature, or two features in a single crypt, changes diagnosis and hence surveillance interval — three years versus none — on a rule nobody has calibrated. The practical rule to take from this: “it came from a guideline” answers the provenance question, not the derivation question. Ask which one you are actually relying on.
The threshold decides whether the AI helped
A cut point does not only categorise patients. When a model estimates a continuous quantity, the threshold applied downstream determines whether the model’s contribution is scored as a benefit at all — and two sources from 2026 show this from opposite directions.
sources/papers/rosbach-2025-automation-bias-tcp.md had 28 pathologists estimate tumour cell percentage unaided and then with an AI, and scored correctness by whether the estimate landed on the right side of the 25% threshold for direct DNA sequencing. Measured continuously, the AI helped: mean absolute deviation from ground truth fell by roughly 2–3 percentage points. Measured across the threshold, among the cases where readers followed the AI against their own prior judgement it produced 38 new errors against 29 corrections — net harm. Same data, same readers, opposite verdicts, and the difference is entirely the cut point. Since nobody acts on a tumour percentage — they act on whether the sample clears the assay minimum — the thresholded view is the one that reaches the patient. Any evaluation of a decision aid for a thresholded marker that reports only the continuous accuracy gain can therefore report a success the decision does not share. See Cognitive Bias in AI-Assisted Diagnosis.
The same paper’s headline 7% automation-bias rate inherits the same dependence: it is defined entirely by that 25% cut, a case estimated at 24% against a truth of 26% counts as a full error, and different assays specify different minimum tumour fractions. The sensitivity of the number to the threshold is not examined.
sources/papers/koyun-2026-slide-selection-uncertainty.md shows the construction step rather than the consequence. Its “high recurrence risk” label does not come from a clinical assay at all: an RNA-derived 21-gene surrogate was prevalence-matched — the cutoff chosen so that about 15% of patients are high risk, to match the proportion scoring ≥26 on the commercial assay in the National Cancer Database — yielding a fixed normalised threshold of 0.7169 on a score spanning −2.009 to 2.744. The paper is careful that this is a research classification threshold and is not numerically equivalent to the clinical cutoff, and that no paired commercial measurements existed to check it against. That care is the point: a prevalence-matched cut point reproduces the base rate of the assay it imitates without reproducing the assay, so every downstream metric is conditional on a matching assumption rather than on a measurement.
Why it matters for my work
The machinery to do this correctly is already in the group’s standard toolchain, so the risk is procedural rather than technical — but this page previously pointed at the wrong module, and that matters for anyone installing a focused module rather than the full one.
Corrected from source on 2026-07-26: OncoPath has four analyses and none of them is cut-point analysis. Categorisation and cut-point selection live in ClinicoPath jamovi Module as optimalcutpoint and categoricaladvanced, and in ClinicoPathDescriptives as categorize. The chapter’s rule still stands — data-derived cut-points overfit unless validated, so use a split-sample or cross-validated option and document the method — it is simply enforced somewhere else.
The issue also surfaces wherever a heterogeneous marker is being quantified. In HER2 Intratumoral Heterogeneity, the value being thresholded itself depends on the sampling scheme, which makes any single cut-point less stable than it appears.
How it connects
ClinicoPath jamovi Module — optimalcutpoint is where the naive and validated cut-point procedures actually are, so choosing correctly costs nothing once the right module is installed.
ClinicoPathDescriptives — categorize covers plain continuous-variable categorisation (tertiles, quartiles) without the threshold search.
Intratumoral Heterogeneity — makes the underlying measurement sampling-dependent, which compounds cut-point instability.
Decision Curve Analysis — reframes threshold choice as an explicit harm-benefit trade-off rather than a statistical search, which is the more defensible framing.
External Validation — the only real test of whether a derived cut-point transfers.
Mitotic Count — the same near-threshold instability in a diagnostic rather than research setting: a systematic measurement bias only changes the answer for cases sitting on a band boundary, which is exactly where it does change it.
Ki-67 Proliferation Index — this page’s canonical dichotomised marker, and the concrete reason guideline thresholds matter: its scoring is irreproducible enough that a national recommendation exists to standardise it.
Aust 2010 — Serrated Polyp Diagnostic Criteria (German consensus) — the counter-example to trusting a guideline threshold on provenance alone: a national consensus threshold with no derivation, no operating characteristics and a three-year surveillance interval hanging on it.
Cognitive Bias in AI-Assisted Diagnosis — where the choice of threshold flips the sign of “did the AI help?” on one dataset, which is this page’s argument made operational.
Rosbach 2025 — Automation Bias in AI-Assisted Tumour Cell Percentage Estimation under Time Pressure — the 25% direct-sequencing cut on tumour cell percentage, and the source of the continuous-versus-thresholded contradiction above.
Koyun 2026 — Slide Selection Introduces Sampling-Induced Prediction Uncertainty — a prevalence-matched threshold built to imitate a commercial assay’s base rate, with the authors’ own warning that it is not the clinical cutoff.
Interobserver Agreement — the other half of the same decision. A threshold and the agreement achievable on the features it counts are not separable questions, and that paper leaves both unmeasured.
Open questions
- Do the group’s existing biomarker papers report how their cut-points were derived? Worth auditing. [unverified]
- Is there a house preference between pre-specified guideline thresholds and internally derived ones? A stated default would prevent per-paper drift.