# Biomarker cutoff selection and validation

Source: https://onco.cc/terms/biomarker-cutpoint-selection/  
OnCo record `biomarker-cutpoint-selection` (Term). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

The threshold used to call a test positive is part of the test's evidence. A cutoff selected because it looked best in one dataset needs evaluation beyond that same dataset.

## Summary

The primary Cutoff Finder study demonstrated methods that optimize biomarker thresholds against distributions, outcomes or survival. Such an optimized split is a discovery result. Its performance must retain the selection procedure and be assessed with appropriate independent evaluation before a clinical rule is claimed. The assay scale and intended decision also matter: a threshold chosen for one drug or purpose is not automatically transferable to another.

## Fields

- Kind: Term
- Last checked: 2026-11-11
- Also known as: biomarker cutpoint selection; data-derived biomarker threshold; optimal biomarker cutoff

## Notes

- A cutpoint applies a category rule to a continuous measurement. Analytical thresholds concern detection or quantitation; diagnostic thresholds classify a defined reference condition; protocol entry cutoffs determine eligibility; treatment-selection thresholds require evidence for an action. Detectable, quantifiable, positive and actionable are separate claims. Categorizing a continuous score does not make biology binary or imply a sharp switch across the line. Nor are cutoffs automatically interchangeable across assays, scales, model versions, drugs or intended uses.
- Keep continuous measurements, reference outcomes, intended action, exact cutoff, units, direction, boundary equality, missing-result handling and assay or model version. Report counts and performance uncertainty. In a fixed fully assessed dataset where higher scores are positive, lowering the cutoff cannot lose an existing positive call, but may find extra cases and add false calls; performance can stay unchanged across an empty score range. Selecting the most favourable outcome-based cutoff can exaggerate separation and discard continuous information. Disclose the search, lock the rule and independently evaluate it in a relevant population for the intended action. Statistical separation alone does not establish clinical benefit.
- Fictional scores for four reference-affected samples are 0.20, 0.55, 0.75 and 0.90; six unaffected samples have 0.10, 0.25, 0.35, 0.45, 0.60 and 0.80. Count equality as positive. At score at least 0.7: TP 2, FP 1, FN 2, TN 5, giving sensitivity 50 per cent, specificity 83.3 per cent and PPV 66.7 per cent. At least 0.4: TP 3, FP 3, FN 1, TN 3, giving sensitivity 75 per cent, specificity 50 per cent and PPV 50 per cent. Lowering this cutoff finds one additional affected sample and creates two additional false calls; sensitivity alone does not choose the better clinical action.
- A mathematical optimization is only as appropriate as its stated objective and error costs. Using the same fictional counts, stipulate cost 5 for a missed case and 1 for a false call: the 0.7 rule costs 11 units and the 0.4 rule 8. Reverse the emphasis to costs 1 and 10, and the totals become 12 and 31. The preferred rule changes without changing any test result. These invented costs are not clinical valuations; actual decisions require relevant benefits, harms and evidence, not a universal mathematical maximum.

## Sources

- Primary biomarker cutoff-optimization methods (2012): https://doi.org/10.1371/journal.pone.0051862
- FDA: reporting results of diagnostic-test studies: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/statistical-guidance-reporting-results-studies-evaluating-diagnostic-tests-guidance-industry-and-fda
- Dichotomizing continuous predictors in multiple regression: a bad idea: https://doi.org/10.1002/sim.2331
- CLSI: harmonized analytical terminology: https://htd.clsi.org/listterms.asp?button=Submit
- FDA: statistical guidance for reporting diagnostic-test studies: https://www.fda.gov/media/71147/download

## Connected records

- terms: [Biomarker](https://onco.cc/terms/biomarker/), [Data leakage in model evaluation](https://onco.cc/terms/data-leakage/), [Evidence from measurement to patient benefit](https://onco.cc/terms/measurement-to-benefit-evidence/), [External validation](https://onco.cc/terms/external-validation/), [Prognostic and predictive biomarkers](https://onco.cc/terms/prognostic-versus-predictive-biomarkers/)

---
JSON: https://onco.cc/api/v1/entities/biomarker-cutpoint-selection.json