Chapter 1

Chapter 1: What Is a Biomarker? From Measurement to Clinical Decision

A number becomes a biomarker only when we know what was measured, how, in whom, and what decision it is meant to support. This chapter builds that vocabulary and shows where each part of WIA-BIO-004 fits.

A laboratory analyser that prints "CRP 12.5 mg/L" has produced a measurement. It has not yet produced clinical knowledge. To turn that number into something a physician can act on, several further facts are needed: the assay that generated it, the specimen it came from, whether the specimen was haemolysed or left on the bench too long, the interval of values seen in comparable healthy people, and, above all, the question the result is supposed to answer. Is it being used to diagnose an acute infection, to monitor an inflammatory disease under treatment, or to estimate long-term cardiovascular risk? The same analyte, measured on the same analyser, plays a different role in each case and must be judged by different evidence.

WIA-BIO-004 is a data standard for exactly this gap between a raw measurement and a defensible clinical claim. It defines how biomarker measurements, reference ranges, validation studies and performance reports are recorded, which statistics are used to judge them, and which protocols and privacy controls surround them. Before we can use the standard, however, we need a precise idea of what a biomarker is and of the two different ways biomarkers are classified. That is the subject of this chapter. The later chapters then follow the life of a biomarker from the statistics of diagnostic performance (Chapter 2) to conformance testing (Chapter 8).

1.1 Definition and the FDA-NIH BEST Vocabulary

The WIA-BIO-004 specification (§1.4) defines a biomarker as a measurable biological characteristic indicating normal processes, disease, or response to therapy. The wording follows the glossary that the specification itself cites as its terminology reference: BEST (Biomarkers, EndpointS, and other Tools), a resource first published in 2016 by a joint working group of the US Food and Drug Administration and the National Institutes of Health. BEST defines a biomarker as a defined characteristic measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention.

Three words in that definition carry most of the weight. Measured means the characteristic is obtained by an instrument or a defined procedure, not reported by the patient; how a person feels, functions or survives is a clinical outcome, not a biomarker. Defined means the characteristic, the method and the unit are specified well enough that another laboratory could obtain the same kind of number. Indicator means the value stands for something else: a biological state, a disease process, or the effect of a drug. A biomarker is therefore always a proxy, and the core scientific question is how good the proxy is for the purpose at hand.

BEST does not classify biomarkers by what they are made of. It classifies them by the question they help answer. There are seven categories. They are listed in Table 1-1 together with a well-established example of each and the consequence that the category has for the data you must record.

Two points about Table 1-1 deserve emphasis. First, the categories are not properties of a molecule. Troponin measured once in the emergency department is a diagnostic biomarker; troponin measured repeatedly in patients receiving a cardiotoxic drug is a safety biomarker. The regulatory term for this is the context of use: a statement of what the biomarker is used for, in which population, and for which decision. Evidence is only valid for the context of use in which it was gathered.

Second, the difference between prognostic and predictive is subtle and is one of the most frequent sources of over-claiming. A marker is prognostic if it separates patients with good and poor outcomes; it is predictive only if the effect of a particular treatment differs between marker-positive and marker-negative patients. A single-arm study in which every patient received the drug cannot distinguish the two: marker-positive patients might do better simply because they had a milder disease. Establishing a predictive claim requires comparison with patients who were not treated, or were treated differently. Chapter 6 returns to this when it discusses clinical correlation and treatment response.

BEST also introduces the idea of a surrogate endpoint: a biomarker that is accepted as a substitute for a clinical outcome in a trial, because a change in the marker reliably predicts a change in how patients feel, function or survive. Blood pressure and LDL cholesterol are familiar examples in cardiovascular medicine. Surrogacy is the strongest claim a biomarker can carry and needs evidence well beyond diagnostic accuracy. WIA-BIO-004 does not define a surrogate-endpoint record, and this book does not treat it further, but it is useful to know where the top of the evidence ladder lies.

1.2 Two Independent Axes: Modality and Clinical Role

Every biomarker can be described along two axes that have nothing to do with each other.

• Modality describes what is measured and by which kind of technology. In WIA-BIO-004 this is the enumeration biomarkerType, with exactly five values: protein, genetic, metabolic, imaging and cellular. The same five values appear in the measurement format (§7.1), in the BiomarkerValidation interface of the API (§8.2) and in the SDK type BiomarkerType.
• Clinical role describes what the result is used for. It corresponds to the BEST categories of Table 1-1. In the standard it appears as the type field of the biomarker definition in the data-format annex (for example "type": "prognostic") and as the type parameter of the registry API in the interface annex ("type": "diagnostic", ?type=predictive).

The specification's own section 2 classifies biomarkers by modality only; the README's list of use cases (diagnosis, drug-response prediction, prognosis, treatment monitoring, risk stratification, clinical trials, population screening) is organised by role. The field name type makes matters worse: in the validation-study record of §7.2, biomarker.type holds a modality ("protein"), while in the biomarker definition the same name holds a role. Because one name carries two meanings, it is easy to mix the axes. A query for "all protein biomarkers" and a query for "all diagnostic biomarkers" return overlapping but different sets, and a system that infers one from the other will make mistakes: PSA is a protein used for both diagnosis and monitoring; EGFR mutation status is genetic and predictive; tumour size is an imaging biomarker that is prognostic in one context and a response marker in another.

This book therefore applies a simple rule. Modality belongs to the measurement; role belongs to the claim. A measurement record says what was measured and how; it is true regardless of how the value is later used. A biomarker definition or validation study says what the value is used for, in which population and with which cutoff; the same measurements may support several such claims. The following fragments show the two places in WIA-BIO-004 JSON. The identifiers and the analyte are fictional and serve only to illustrate structure.

// Measurement record (§7.1): modality lives here
{
  "standard": "WIA-BIO-004",
  "version": "1.0.0",
  "measurement": {
    "biomarkerID": "BM-EXAMPLE-01",
    "biomarkerName": "Example serum protein X",
    "biomarkerType": "protein",
    "value": 4.2,
    "unit": "ng/mL",
    "sampleType": "blood_serum"
  }
}

// Biomarker definition (data-format annex): role lives here
{
  "biomarkerId": "BM-EXAMPLE-01",
  "name": "Example serum protein X",
  "type": "diagnostic",
  "indication": { "disease": "Example condition", "population": "Adults with suspected disease" }
}

The measurement record contains no role, and the definition contains no individual values. When the same protein is later evaluated as a monitoring marker, a second definition with "type": "monitoring" and its own validation evidence is created; the measurement records do not change. Chapter 3 describes both record types in full, and Chapter 8 lists the ambiguity of the type field among the open issues of version 1.0.

1.3 Molecular Biomarkers

Three of the five modalities are molecular: proteins, nucleic acids and small-molecule metabolites. They share a specimen-based workflow (a sample is collected, processed, stored and analysed), which is why the pre-analytical requirements of Chapter 5 apply to all of them. They differ in what kind of value the assay returns and in the errors that typically creep into the data.

Protein biomarkers

The specification lists PSA (prostate-specific antigen), troponin, C-reactive protein (CRP) and HbA1c as examples, measured by ELISA and other immunoassays, mass spectrometry, immunohistochemistry and Western blotting. Most routine protein biomarkers are measured by immunoassay: an antibody binds the target and a signal proportional to the amount bound is read out. This has an important data consequence. Different manufacturers use different antibodies that recognise different parts, or different forms, of the same protein, and their calibrators may not be traceable to a common reference material. Two immunoassays for "the same" analyte can therefore report numerically different values for the same specimen. That is why the measurement record carries a measurementMethod field and a laboratoryID: without them, values from different sites cannot be pooled safely.

HbA1c illustrates a second problem, unit heterogeneity. It is a glycated form of haemoglobin that reflects average blood glucose over the lifespan of red blood cells (roughly the preceding two to three months). It is reported either as a percentage of total haemoglobin (the NGSP convention) or in mmol/mol (the IFCC convention). Both are legitimate, but a value of "48" is meaningless unless the unit is stated. Units matter even within one convention: CRP is often reported in mg/L but sometimes in mg/dL, a factor of ten apart.

Genetic biomarkers

Genetic biomarkers are based on DNA or RNA. The specification's examples are BRCA1/2 mutations (hereditary breast and ovarian cancer risk), EGFR mutations (targeted therapy in lung cancer), the BCR-ABL fusion (chronic myeloid leukaemia) and HLA alleles (autoimmune disease risk), measured by PCR, next-generation sequencing (NGS), microarrays, and quantitative reverse-transcription PCR (qRT-PCR) for gene expression.

The results are often categorical rather than continuous: a variant is present or absent, an allele is carried or not. Some are quantitative, such as the ratio of BCR-ABL transcripts to a control gene used to monitor treatment, or the variant allele fraction reported by sequencing. Three data details are easy to lose. The first is the distinction between germline variants, inherited and present in every cell, and somatic variants, acquired by a tumour; the same gene can carry either, with very different meaning. The second is the reference genome build against which positions are reported, since coordinates differ between builds. The third is the variant nomenclature, for which the HGVS recommendations are the common convention. Chapter 3 shows how genetic results travel in VCF files and how they are linked to the WIA-BIO-004 records.

Genetic results also raise the strongest privacy concerns in the field. A rare variant can identify a person, and it reveals information about blood relatives who never consented to testing. Chapter 7 treats rare-variant suppression and tiered consent in detail.

Metabolic biomarkers

Metabolic biomarkers are small molecules: glucose (diabetes), cholesterol and other lipids (cardiovascular disease), creatinine (kidney function) and lactate (tissue hypoxia) in the specification's list. They are measured on clinical chemistry analysers, by liquid chromatography with tandem mass spectrometry (LC-MS/MS), or by nuclear magnetic resonance (NMR) spectroscopy.

Small molecules are often the most sensitive to what happens before the analysis. Glucose keeps falling in a blood sample that has not been separated or treated with a glycolysis inhibitor, because blood cells continue to consume it; lactate, a product of the same glycolysis, rises. Lipids, especially triglycerides, depend on fasting status. This is why the measurement record has room for fastingStatus and why the time between dateCollected and dateMeasured matters. Metabolic values also come with competing unit systems: glucose, for instance, is reported in mg/dL in some countries and in mmol/L in others, and converting between them requires the molar mass of glucose (about 180 g/mol, so 1 mmol/L corresponds to about 18 mg/dL). Some important metabolic markers are not measured directly at all: an estimated glomerular filtration rate is calculated from creatinine with an equation, and the record must state which equation.

1.4 Imaging and Cellular Biomarkers

Imaging biomarkers

Imaging biomarkers are quantitative features extracted from medical images: tumour size (cancer progression), ejection fraction (cardiac function), bone density (osteoporosis) and brain atrophy (neurodegeneration) are the specification's examples, obtained with CT, MRI, PET and ultrasound. In clinical practice bone mineral density is most often measured by dual-energy X-ray absorptiometry (DXA), which the specification's list of modalities does not name; the data principles below apply regardless.

An imaging biomarker is always a derived value. Between the patient and the number lie an acquisition protocol, a scanner and its settings, a reconstruction algorithm, and a reader or a segmentation program. Change any of them and the value can change. Left-ventricular ejection fraction measured by echocardiography and by cardiac MRI is the same concept but not the same measurement; the diameter of a tumour depends on which lesion was selected and in which plane it was measured. For tumour size in treatment trials, the RECIST 1.1 criteria define which lesions to measure and how to sum them. For the data record, this means that measurementMethod should name the modality, the protocol and the software version, and that inter-reader variability must be treated as part of analytical validation (Chapter 5).

There is also a structural gap in version 1.0. The sampleType enumeration has eight values, all of them biological specimens (blood_serum, blood_plasma, whole_blood, urine, tissue_biopsy, cerebrospinal_fluid, saliva, stool). None fits an image. Until the standard adds one, implementers should leave the field out for imaging records rather than force a wrong value, and describe the acquisition in measurementMethod. This gap is recorded again among the open issues in Chapter 8.

Cellular biomarkers

Cellular biomarkers count or characterise cells: CD4 T-cell count (HIV monitoring), circulating tumour cells (metastatic cancer), neutrophil count (infection and inflammation) and platelet count (clotting disorders). They are measured by flow cytometry, cell sorting, microscopy and automated cell counters.

Cell counts are usually reported as absolute concentrations (for example cells per microlitre), but some are also reported as percentages of a parent population; CD4 is commonly given both ways, and the two are not interchangeable. Flow-cytometric results depend on the gating strategy, the choice of which events are counted as the target population, and on the age of the specimen, because cells deteriorate in stored whole blood. Rare-cell assays such as circulating tumour cell counts work near the limit of detection, where a difference of one or two cells can change a classification; the limit-of-detection methods of Chapter 5 are central here.

The "typical clinical roles" column is a guide, not a rule. Each role must be established by its own evidence in its own context of use, which is exactly what the rest of the book is about.

1.5 The Biomarker Life Cycle

Biomarkers do not arrive in clinical practice fully formed. The quality-metrics annex of WIA-BIO-004 describes three validation levels that follow an initial discovery stage. Together they form the life cycle that organises this book.

1. Discovery. Many candidate markers are screened, often in small groups of cases and controls, with the aim of finding signals worth pursuing. The danger at this stage is false discovery: when hundreds of candidates are tested, some will look promising purely by chance. Discovery results are hypotheses, not claims. The SDK's ValidationStudy type records this stage with studyType: 'discovery'.
2. Analytical validation. Does the assay measure the analyte reliably? This covers limit of detection and quantification, precision, accuracy, linearity and stability. The annex's guidance is on the order of 20 to 50 measurements per parameter. Chapter 5 describes the protocols.
3. Clinical validation. Does the measured value separate the relevant clinical states in the intended-use population? This is where sensitivity, specificity, predictive values and ROC-AUC are estimated. The annex's guidance is 200 to 500 or more participants. Chapter 2 develops the statistics and Chapter 6 the study design.
4. Clinical utility. Does using the biomarker change decisions and improve outcomes compared with not using it? Here the annex speaks of 500 to 1,000 or more participants and of "improved patient management" rather than a single accuracy number. Chapter 6 covers reclassification and decision-curve methods.

The sample sizes above are the standard's own guidance figures. They are not a substitute for a sample-size calculation for a specific study, which Chapter 6 explains, and the acceptance criteria attached to each level in the annex differ from those in other WIA-BIO-004 documents; Chapter 5 reconciles them.

   DISCOVERY  ──►  ANALYTICAL      ──►  CLINICAL          ──►  CLINICAL
   candidates      VALIDATION           VALIDATION             UTILITY
                   "does the assay      "does the value        "does using it
                    measure well?"       separate states?"      improve care?"
       │                 │                     │                      │
       │            Chapter 5            Chapters 2, 6            Chapter 6
       │         LOD, LOQ, CV,         Sens, Spec, PPV, NPV,    NRI, IDI,
       │         recovery, stability   LR, ROC-AUC, cohorts     net benefit
       │                 │                     │                      │
       └───────◄─── failure or assay change sends the marker back ◄──┘

   ═══════════════ runs underneath every stage ═══════════════
   Chapter 3  data formats (measurement, reference range, study, report)
   Chapter 4  SDK, CLI and REST interfaces
   Chapter 7  integration, security and privacy
   Chapter 8  conformance testing and evidence

Two features of Figure 1-1 are easy to overlook. The first is the feedback arrow. The stages are not a one-way street: a marker that fails clinical validation often goes back to discovery, and any change to the assay (a new antibody lot, a new platform, a new laboratory) reopens analytical validation and may require clinical results to be confirmed again. A data standard must therefore record which assay version produced each result, so that evidence can be traced to the method that generated it.

The second feature is that analytical and clinical validation answer different questions and cannot substitute for each other. An assay can be perfectly precise and still useless clinically, because the analyte simply does not differ between diseased and healthy people. Conversely, a marker with a strong biological link to disease will look weak if the assay is imprecise, because measurement noise blurs the separation between groups. Clinical performance is always bounded by analytical performance, which is why Chapter 5 comes before the clinical validation chapter in the order of work, even though Chapter 2 introduces the clinical statistics first so that readers know what the analytical work is for.

1.6 Why Biomarker Data Needs a Standard

Many biomarkers that look promising in their first publications perform less well when others try to reproduce them. Part of the reason is biological, but a large part lies in how the data were produced and reported. Five sources of irreproducibility recur, and each of them corresponds to something WIA-BIO-004 asks implementers to record or control.

Pre-analytical variation

Everything that happens to a specimen before it reaches the analyser can change the result: whether the patient was fasting, the time of day (many analytes follow a circadian rhythm), posture during collection, tourniquet time, the delay before processing, centrifugation, storage temperature and the number of freeze-thaw cycles. The specification (§3.1.2) lists these factors and sets a few limits, such as a tourniquet time under one minute, processing of plasma within two hours and fewer than three freeze-thaw cycles. Pre-analytical variation is invisible in the final number unless it is written down alongside it.

Unit and assay heterogeneity

As Section 1.3 showed, different assays for the same analyte can disagree, and the same quantity can be expressed in different units. A small calculation shows how dangerous an implicit unit is. The CRP example of §7.1 is 12.5 mg/L, above the record's reference range of 0 to 10 mg/L. If a receiving system dropped the unit and assumed mg/dL, it would read 12.5 mg/dL, which equals 125 mg/L: a tenfold overstatement that turns a mildly raised value into a very high one. Chapter 3 shows how controlled codes (LOINC for the test, UCUM for the unit) remove this ambiguity.

Unreported or data-driven cutoffs

A continuous biomarker becomes a positive or negative test only at a cutoff. If the cutoff is chosen after looking at the data, typically the value that maximises accuracy in the same sample, the reported sensitivity and specificity are optimistic, because the cutoff has been fitted to that sample's random fluctuations. If the cutoff is not reported at all, the result cannot be reproduced. The WIA-BIO-004 validation-study record therefore carries the chosen cutoff (optimalThreshold, with its value, unit and Youden index), and the rule used to choose it should be stated alongside; Chapter 2 compares the rules.

Spectrum bias

Accuracy depends on who is tested. A study that compares patients with advanced, obvious disease to young, healthy volunteers will find large differences, because the two groups are far apart. In real use the marker must distinguish early or mild disease from other conditions that resemble it, and the separation is smaller. Estimates from such "extreme" case-control designs are usually too optimistic for the intended-use population. The validation-study record therefore describes the population (numbers of cases and controls, age range, sex and ancestry distribution) rather than only the totals, and a study report should add the inclusion criteria and clinical setting.

Non-blinded measurement and the reference standard

If the person measuring or reading the biomarker knows the patient's clinical status, the measurement can drift towards the expected answer; this matters most where human judgement is involved, as in imaging or microscopy. A related problem arises when the biomarker itself forms part of the reference standard used to define who has the disease, which inflates apparent accuracy. Blinding and an independent reference standard are among the validation requirements below. For reporting diagnostic accuracy studies more generally, the STARD guideline is the widely used checklist.

The seven validation requirements

The WIA-BIO-004 README summarises the evidence it expects before a biomarker is considered validated. They are previewed here because every later chapter refers back to them.

1. Sample size: a minimum of 100 cases and 100 controls for initial validation.
2. Statistical power: at least 80 % power to detect clinically meaningful differences.
3. Multiple cohorts: validation in at least two independent populations.
4. Blinding: measurements performed without knowledge of clinical status.
5. Reference standards: comparison against a gold standard for diagnostic biomarkers.
6. Reproducibility: a coefficient of variation below 15 % for quantitative biomarkers.
7. Stability: documented pre-analytical stability.

Two of these requirements are stated differently in other parts of the standard, and readers should know this from the start. The README asks for 100 cases and 100 controls, whereas the SDK constant MIN_VALIDATION_SAMPLES is 100 measurements in total; Chapter 4 explains what the SDK actually checks. The 15 % CV limit is one of several precision limits found across the specification, the quality-metrics annex and the SDK, which distinguish diagnostic, screening and research use; Chapter 5 sets them side by side. Throughout this book, whenever a numeric limit is quoted, the document it comes from is named.

1.7 Map of WIA-BIO-004

WIA-BIO-004 is not a single document. It consists of a main specification, four phase documents that expand particular topics, a TypeScript SDK and a command-line tool. Table 1-4 lists the parts and the chapters in which they are used.

Note that the file names of phases 3 and 4 do not match their content: the file called "PROTOCOL" contains the quality metrics, and the file called "INTEGRATION" contains security and privacy. The phase documents also carry an older identifier (WIA-BIO-003) in their headers. This book treats all four as the phase annexes of WIA-BIO-004 and uses that identifier throughout; the naming inconsistency is listed among the open issues in Chapter 8.

Section 8.1 of the main specification states what any WIA-BIO-004 compliant system must include. The five required components give a compact picture of the whole standard:

1. Data collection: standardised capture of biomarker measurements (Chapter 3).
2. Quality control: automated QC checks and validation (Chapter 5).
3. Statistical analysis: ROC, sensitivity and specificity calculations (Chapters 2 and 4).
4. Clinical correlation: integration with patient outcomes (Chapter 6).
5. Reporting: standardised performance reports (Chapters 3 and 4).

Chapter 7 shows how these components are connected in a working system, and Chapter 8 explains how to demonstrate that a system actually has them. One caution about the tools, discussed fully in Chapter 4: the version 1.0 command-line tool prints fixed demonstration output and does not yet read the data file it is given. Real calculations should be done with the SDK.

Reading paths

The chapters build on each other, but not every reader needs them in order. Three paths cover the most common roles.

Summary

A biomarker is a defined, measured characteristic that indicates a biological process, a disease process or a response to an intervention. What it is made of (its modality) and what it is used for (its clinical role) are separate questions. WIA-BIO-004 records modality in the measurement as biomarkerType, and role in the biomarker definition and validation study; this book keeps them apart. Each modality has typical data pitfalls: assay dependence for proteins, germline/somatic and genome-build details for genetic markers, pre-analytical instability for metabolites, protocol and reader dependence for imaging, and gating and reporting conventions for cellular markers. A biomarker moves through discovery, analytical validation, clinical validation and clinical utility, and the evidence from one stage does not replace another. Much of the irreproducibility in the field comes from unrecorded pre-analytical conditions, unit and assay differences, undeclared cutoffs, spectrum bias and unblinded measurement, and the fields and requirements of WIA-BIO-004 exist to make these visible.