A measurable signal can be useful without answering every clinical question. Biomarker research asks what a measurement represents and how it performs in a stated setting. Read the proposed use before deciding whether a study supports diagnosis, monitoring or a prediction. This is a guide to research terminology, not an interpretation of a personal test result.
A biomarker needs a context of use
A biomarker is a measured characteristic used as an indicator of a biological process or response. Biomarker discovery identifies potentially useful candidates. Biomarker validation then asks whether a candidate meets the requirements for a particular intended use; an interesting association alone does not settle that question.
Write the proposed use as a sentence: what is measured, in whom, by which method, and to inform which question? A signal associated with a condition in one population may not have the same performance in another. Do not extend the setting beyond the study without supporting evidence.
Diagnostic, monitoring and predictive questions
Diagnostic biomarkers help identify a condition or subtype in a defined context. Monitoring biomarkers are measured repeatedly to assess change. Prognostic biomarkers concern the likelihood of a future outcome. Predictive biomarkers concern the likelihood of a response to a particular intervention. The FDA–NIH BEST resource distinguishes these uses.
The categories describe questions, not a universal quality ranking. A measurement may have more than one proposed use, but each needs its own evidence. A result showing that a marker is associated with worse outcomes does not, on its own, show which intervention would work best for people with that result.
Analytical performance and clinical usefulness
Analytical validation examines whether an assay measures its intended characteristic reliably under specified conditions. Laboratory quality control checks ongoing performance; quality assurance describes the wider processes used to support reliable work. Clinical usefulness is a separate question about whether using the result improves an intended decision or outcome.
When reading an assay paper, locate the material tested, measurement range, repeatability and relevant sources of interference. These headings are a reading framework, not a substitute for a validated laboratory protocol or accreditation standard. A technically repeatable result can still be unsuitable for a proposed clinical decision.
Sensitivity, specificity and the tested population
Sensitivity concerns the proportion of people with the target condition who test positive; specificity concerns the proportion without it who test negative. Predictive values ask a different question: what does a positive or negative result mean in the population being tested? The prevalence and selection of that population matter.
In an illustrative set of 100 people known to have a target condition, a test identifying 90 has 90% sensitivity in that set. That is not a claim that 90% of positive results in every population are correct. Check how the reference condition was established and whether uncertain results were excluded before accepting a headline percentage.
Reference ranges and result interpretation
Reference ranges describe the distribution of results in a defined reference population, using a stated method. They are not automatically identical to decision thresholds or a complete definition of health. Laboratory result interpretation brings together the result, method and relevant context; a paper's population-level finding should not be substituted for an individual's clinical assessment.
Retain units and the laboratory's stated interval. Compare like with like before drawing conclusions from two reports. If a study presents a new cut-off, find the independent evaluation rather than assuming performance in the development dataset will be repeated elsewhere.
Questions to carry into the full paper
Identify who was included, who was left out, the reference method, missing results and the planned analysis. Keep development and independent validation datasets separate. A promising first report can justify further research without justifying routine use.
Read clinical research designs for comparison groups and outcomes, and prevention and screening evidence for the additional questions raised when testing people without symptoms. Clova's journal does not provide a diagnostic testing service.
Sources and further reading
Sources checked 25 September 2026: FDA–NIH BEST resource; NIEHS: biomarkers; UK NSC: screening benefits and harms.
Calculate sensitivity and specificity from the counts
Diagnostic sensitivity describes the share of reference-positive people whom the index test identifies as positive. Specificity describes the share of reference-negative people whom it identifies as negative. The reference standard is the comparison used to establish the target condition; it is not automatically infallible. The NICE review-methods manual explains why the population, index test, reference standard and accuracy outcomes belong in the research question. These definitions do not establish an acceptable threshold for a particular clinical use.
Here is an invented arithmetic exercise, not patient data or a result for a real test. Of 100 people with the reference condition, a test returns 90 positives and 10 negatives. Of 900 without it, the test returns 45 positives and 855 negatives. The four cells are therefore 90 true positives, 10 false negatives, 45 false positives and 855 true negatives. All 1,000 observations are accounted for.
Sensitivity is 90 divided by (90 + 10), or 90%. Specificity is 855 divided by (855 + 45), or 95%. Overall classification accuracy is (90 + 855) divided by 1,000, or 94.5%. These three percentages use different denominators. A report that gives only 94.5% accuracy conceals both the missed reference-positive cases and the false alarms.
Why a positive result is not the same as sensitivity
In that exercise, positive predictive value is 90 divided by (90 + 45), or about 66.7%. Negative predictive value is 855 divided by (855 + 10), or about 98.8%. The first asks how many positive results are reference-positive; the second asks how many negative results are reference-negative. Neither is interchangeable with sensitivity or specificity.
Now change the fictional population to 10,000 people, including 100 with the reference condition. Hold sensitivity at 90% and specificity at 95% purely to demonstrate the arithmetic. The resulting cells are 90 true positives, 10 false negatives, 495 false positives and 9,405 true negatives. Positive predictive value becomes 90 / 585, about 15.4%, despite unchanged assumed sensitivity and specificity. This is a constructed example, not a prediction for a screening programme.
The calculation shows why a positive result needs its population context. In real research, the spectrum of people tested can also change sensitivity and specificity; holding them constant is an assumption of the example. The BMJ methods paper on the spectrum effect explains this additional issue. Do not transfer a hospital estimate to an asymptomatic population without evaluating that difference.
Interpret an accuracy estimate and its uncertainty
Keep the four counts, the positivity threshold and the reference definition beside every estimate. A rate based on 9 successes out of 10 and a rate based on 900 out of 1,000 both equal 90%, but their statistical precision differs. A confidence interval reports uncertainty under a specified estimation method; it does not correct selective recruitment, a flawed reference standard or an omitted group.
When two methods are applied to the same people, the observations are paired. Simply treating the two sets of results as unrelated loses information about who received conflicting classifications. Before reading a claimed improvement, identify the paired comparison, its uncertainty and whether the operating threshold was chosen before the evaluation. A percentage without that context is not enough to declare one test superior.
Missing results, thresholds and misleading comparisons
An original audit exercise: 120 people were enrolled, but an accuracy table contains 100. Of the other 20, eight had an unreadable index test and twelve lacked the reference assessment. A percentage calculated from the table describes the analysed 100, not necessarily all 120 who entered. Ask for the flow of participants and the handling of both missing groups; do not silently classify them as correct results.
Choosing the threshold that looks best in a development dataset and assessing it again in those same records can make performance look optimistic. The STARD 2015 explanation distinguishes planned and exploratory thresholds, discusses indeterminate or missing results, and explains the reporting of participant flow and precision. Its reporting guidance helps a reader inspect a study; completing a checklist does not by itself validate a test.
A diagnostic-result comparison you can reproduce
Use one record per study and operating threshold. Enter the target condition, population, recruitment setting, index test, reference standard, four counts, excluded observations, estimate and interval. Recalculate the denominators before copying a headline. Keep a second record for a genuinely different threshold instead of mixing one sensitivity with another specificity.
For the fictional 1,000-person example, the record should reproduce 90%, 95%, 66.7% and 98.8% for sensitivity, specificity, positive predictive value and negative predictive value. If it does not, inspect the cell labels and the denominator. This is a worked interpretation aid, not a validated laboratory protocol, clinical decision tool or evidence that Clova offers diagnostic testing. No new test-performance result is claimed.
Define the decision before choosing the marker
Biomarker selection starts with a decision, rather than a list of molecules. Specify whether the result will identify a research subgroup, follow change over time or estimate an outcome. Then state the specimen, measurement method, population and time horizon. These choices determine what validation evidence is relevant. The FDA context-of-use explanation provides a useful vocabulary for drug-development research; it does not establish UK authorisation for a test.
Consider an invented study that records a marker at enrolment and follows participants for one year. If high values are associated with an outcome, the observation may support a prognostic question. It does not establish that those participants respond differently to an intervention. That second claim needs a comparison of intervention effects across marker groups. Following a marker repeatedly is yet another design, concerned with monitoring change.
Make a short selection record with the proposed decision, measurement and evidence needed to change that decision. Reject an otherwise impressive validation result when it concerns a different use. A repeatable measurement is valuable, but repeatability alone cannot fill the gap between a risk association and an intervention-selection claim.
Trace a digital measurement through its processing
A digital biomarker is often several steps removed from its sensor. A device captures a physical signal; software processes the observations; another rule may turn them into a summary measure. The V3 framework separates verification of the technology, analytical validation of the measurement and clinical validation for the proposed context. V3 framework for digital measurement validation.
For a fictional activity study, distinguish three questions: does the sensor record movement consistently, does the algorithm recognise the intended movements, and does the resulting summary answer the study question in its participants? Passing the first question cannot answer the other two. A software update may change the second step while leaving the hardware untouched, so the version belongs in the research record.
Compare performance across the conditions actually studied, including missing wear time and unsuccessful recordings. Reporting only successful measurements can conceal the proportion of people for whom no result was available. Keep the sensor, firmware, algorithm, reference measurement and participant group together when comparing two papers. These checks support critical reading; they are not a certification of any particular wearable or app.
Check the specimen record before comparing results
A specimen carries a history before it reaches an analyser. Collection, processing, storage and accompanying information can affect what researchers can learn from it. The NCI biospecimen overview explains why the material and its associated data both matter. A specimen label alone does not establish that two research measurements are comparable.
Suppose an invented comparison uses one group processed immediately and another processed after a long delay. If group and handling method coincide, a difference cannot confidently be assigned to biology alone. Adding more readings from the same stored material would improve neither the original handling comparison nor the independence of the specimens.
In a reading record, retain specimen type, collection timing, processing history, storage information, assay batch and the rules for rejecting unsuitable material. Mark unreported details as unknown. Do not fill them with a preferred laboratory procedure. This makes a specific limitation visible to the next reader and prevents a later summary from treating an undocumented handling step as controlled. For an actual collection procedure, use the study's approved protocol and relevant laboratory documentation.
When can a surrogate biomarker stand in for an outcome?
A surrogate endpoint substitutes for a direct measure of how people feel, function or survive. An association with that outcome is only a starting point: researchers need evidence that an intervention's effect on the proposed surrogate predicts an effect on the specific outcome. The FDA surrogate-endpoint resource distinguishes candidate, reasonably likely and validated surrogates. These are US development categories, not a statement of UK approval.
Imagine two fictional interventions that produce the same change in a laboratory marker. If their effects on the outcome differ, the marker change cannot, by itself, justify assigning them the same clinical benefit. A marker can track one pathway while other effects of the interventions influence the outcome.
A useful reading record names the proposed surrogate, the outcome it replaces, the intervention class, population and evidence supporting that replacement. Preserve the authors' validation category rather than shortening every reference to a biomarker into a claim of a validated endpoint.
This distinction is especially relevant when an abstract highlights a rapid laboratory change while follow-up for the main outcome remains incomplete. Report the measured change and the remaining outcome uncertainty separately.
Could an immunoassay difference come from interference?
An immunoassay result depends on the measurement system as well as the substance being measured. Biotin provides a documented example: in susceptible assays it can distort the signal, with the direction depending on the test's design. The FDA guidance on biotin interference describes this analytical problem. Its existence does not mean that every assay uses the same technology or has the same susceptibility.
Consider a fictional comparison in which one study uses an assay containing a biotin-dependent step and another uses a different design. A disagreement between their reported concentrations raises a question about measurement comparability; it does not establish which result is correct or show a biological change.
For critical reading, record the assay name and version, specimen matrix, stated interference testing and how unexpected results were investigated. Check whether the paper distinguishes an interference experiment from ordinary precision measurements. Repeated measurements can agree closely while sharing the same systematic interference.
Keep conclusions limited to the tested concentrations and conditions. A report that examined one possible interferent has not excluded every source of bias. This is a way to assess a published method, not instructions to change supplements, medicines or individual clinical testing.
Why is a PCR cycle threshold not a universal concentration?
A PCR cycle threshold records when an amplification signal crosses the assay's specified threshold. Turning it into a concentration requires a validated quantitative relationship, rather than assuming that the same number means the same amount in every laboratory. Public Health England's Ct guidance stresses the importance of assured linearity, detection limits and standard quantification curves. That document concerns SARS-CoV-2 assays; its numerical interpretation cannot simply be transferred to other targets.
In an invented research comparison, two reports both show a Ct of 28. They use different target sequences, extraction methods and instruments. The matching number is insufficient evidence that the starting specimens contained identical target concentrations. A study using calibrated copies per unit volume answers a more explicit quantity question.
When reading a PCR diagnostic report, retain the target, specimen type, extraction method, assay and calibration information. Also distinguish a failed or inhibited measurement from a reported target-negative result: the controls and the laboratory's interpretation rules matter.
Without a documented conversion, keep the result in its reported form. Do not turn a Ct alone into a claim about infectiousness, illness severity or a personal decision.
What makes an AI diagnostic test set independent?
An AI model can appear accurate when its evaluation data resemble information already used to build it. Independent testing asks how the final system performs on observations kept separate from training and tuning. The joint FDA, Health Canada and MHRA machine-learning principles identify both independent test sets and representation of the intended population as development concerns.
Consider a fictional imaging dataset containing several scans from each person. Randomly dividing individual scans can leave one person's images in both training and testing. Those test images are not equivalent to entirely new people, even if the filenames differ. The study therefore needs to explain the unit used to separate its datasets.
A second hospital can provide a valuable external evaluation, but the label external does not guarantee that every intended setting was examined. Retain the model version, separation rule, recruitment setting, reference labels and subgroup results alongside the overall performance estimate.
Check whether investigators changed the model or threshold after seeing test results. Such changes create another development step and make a later untouched evaluation relevant. These questions assess a reported AI-assisted diagnostic study; they do not certify a product or establish that an algorithm improves outcomes.
Does detecting an analyte mean its amount can be reported precisely?
Detection and quantification answer different questions. An assay may distinguish a signal from background before it can measure the amount with suitable accuracy and precision. The ICH analytical-validation guideline separates the detection limit from the quantitation limit. These are method characteristics, not universal concentrations that apply to every instrument or specimen.
Imagine a report with a detection limit of 1 unit and a quantitation limit of 3 units. A signal corresponding approximately to 2 units may support detection under the validated conditions, while remaining outside the range for a dependable numerical amount. Writing exactly 2.00 units would imply a level of quantitative confidence the example has not established. Equally, below detection is not the same statement as a proven absence of the analyte.
For analytical validation, record what each limit means, how it was established, the specimen material and the intended reporting range. Check how results between the limits are labelled and how the analysis handles them. Substituting zero for every such result changes the dataset and needs justification. This distinction helps interpret laboratory reporting; it does not establish a clinical threshold or turn a detectable signal into a diagnosis.
Does passing an external quality assessment replace routine laboratory controls?
External comparison and routine control address different parts of laboratory quality. An external assessment examines performance using an independently organised comparison. Internal controls monitor how a procedure is behaving during routine work. The WHO quality-assessment manual places these within a wider system that also includes improvement of the testing process. One successful exercise is not evidence that every later result is accurate.
Consider a fictional laboratory that performs well in an external comparison in March. In June, its routine control values begin to drift after a change of reagent lot. The March result cannot resolve that new signal. Conversely, stable internal readings do not establish agreement with other laboratories if the local procedure has a consistent bias. The two records complement each other rather than allowing one to cancel the other.
For laboratory quality assurance, note the analyte, material, method, time period, acceptance rules and any investigation of unexpected findings. Also distinguish participation in an external programme from satisfactory performance within it. Neither a certificate nor a single control chart evaluates every step from specimen identification to reporting. A useful research summary states which quality question was checked and what remained outside that check.
Is every cell-free DNA measurement a measurement of tumour DNA?
Cell-free DNA names material found outside cells; circulating tumour DNA specifies a tumour origin. The National Cancer Institute definition describes fragments released into blood by tumour cells. A measurement of all extracellular DNA therefore needs additional evidence before it can be described as tumour-derived. Total concentration and a tumour-associated signal are different reported quantities.
In an invented mixture, a specimen contains 100 units of measured cell-free DNA, of which 5 units are attributable to the signal being investigated. A second specimen contains 200 total units and the same 5 attributable units. The total doubled while the attributable amount stayed constant; its proportion halved. This arithmetic illustrates why a concentration, an absolute amount and a fraction must not be interchanged in a liquid-biopsy summary.
Keep the specimen type, collection and processing record, assay target and denominator with the result. Check what supports the claimed origin and whether the method measures a selected sequence, a broad concentration or another feature. A negative result means the reported target was not detected under those conditions; it does not establish the absence of every possible tumour signal. Individual diagnostic conclusions require the relevant clinical investigation, beyond this measurement example.
Supporting references: NCI: cell-free DNA fragmentation research.
Do identical radiomic feature names guarantee identical calculations?
A shared label is not enough. A radiomic feature is calculated through a chain of image-processing decisions, including the region analysed and how image values are represented. The Image Biomarker Standardisation Initiative provides definitions and reference comparisons intended to make those calculations reproducible. Agreement about a feature's name does not document the complete processing chain.
Imagine two teams analysing the same image. One includes a boundary in its region of interest that the other excludes, or their software groups intensities into different bins. They can obtain different values while both report a feature called texture. Comparing the final numbers without the settings cannot identify whether the difference comes from biology, region selection or computation. A successful benchmark comparison checks a specified calculation; it does not validate every later research claim made with it.
Retain the acquisition details, segmentation, resampling, intensity handling, feature definition and software version. Where available, record which standard reference tests were actually passed. Reproducible extraction and prediction of an outcome are separate achievements: a perfectly repeated number can still be uninformative for the proposed outcome. This distinction helps assess imaging biomarkers without treating a software label or a benchmark result as a clinical endorsement.
Can a urinary biomarker concentration change without a matching change in the amount excreted?
A concentration is an amount divided by a volume, so dilution can change it. A timed excretion amount and a concentration from one specimen are not interchangeable measurements. The NIDDK explanation of urine albumin testing also describes relating albumin to creatinine when estimating excretion from a urine sample. That ratio has a different denominator from a concentration alone.
For a purely arithmetic example, 10 milligrams present in one litre gives 10 milligrams per litre. The same amount in two litres gives 5 milligrams per litre. The concentration halves without the total amount changing. This invented example does not describe an individual or explain the cause of a real result; it shows why retaining volume and collection duration matters.
When reading urinary-biomarker research, distinguish a spot concentration, a ratio and a timed collection. Keep their units, collection conditions and calculation methods visible. A ratio can address a dilution question while introducing its own reference measurement; it should not be treated as a direct count of everything excreted. Comparisons across papers need compatible definitions before their numerical differences can be interpreted. Clinical thresholds and personal results require the appropriate clinical source and context.
Top comments (0)