Units Harmonization

Units harmonization is the process of putting measurements from different assays, instruments, labs, and time points onto a scale where a difference in the number means a difference in the thing being measured. It covers three separable problems: unit conversion (mg/dL to mmol/L), scale calibration (two immunoassays reporting the same analyte in pg/mL that disagree by 40%), and semantic alignment (deciding that “Vitamin D, 25-OH, total” from one lab and “25(OH)D” from another are the same test).

Harmonization and standardization are not synonyms. Standardization means traceability to a reference measurement procedure or certified reference material: creatinine is standardized to IDMS, so a value from any compliant lab is metrologically anchored. Harmonization is what you do when no such reference exists, aligning methods to each other so results agree, often via a commutable common calibrator. The COVID-19 anti-spike assays are the clean example: dozens of platforms with incompatible arbitrary units were pulled toward a common scale using the WHO International Standard, reported in BAU/mL¹. Recent work in clinical chemistry argues that measurement units themselves, not just the numbers, need governance, since inconsistent unit reporting is a source of error independent of assay quality².

How it works

A working harmonization pipeline has four stages.

1. Identity resolution. Map each incoming test to a canonical concept. LOINC codes are the right target for blood chemistry because the code encodes analyte, property, timing, system, and scale. Do not map on test name strings. “Ferritin” and “Ferritin, serum” are fine, but “Free T4” and “T4” are different analytes.

2. Dimensional conversion. Once you know the analyte you know its molar mass, so mg/dL to mmol/L is a fixed factor: glucose ÷ 18.016, total cholesterol ÷ 38.67, creatinine × 88.42 to µmol/L. Store the factor with the concept, not in the conversion code. Refuse to convert when the dimension does not match (mass/volume to activity units) rather than guessing.

3. Scale alignment. Where two methods measure the same analyte on different scales, you need either a shared calibrator or a method-comparison regression (Passing-Bablok or Deming, never ordinary least squares, since both axes carry error). This is where most of the work lives. For Alzheimer’s blood biomarkers, cross-platform p-tau217 values are not interchangeable without this step, and harmonized reporting conventions are still being negotiated³. Protease activity assays have a similar problem: activity depends on substrate, buffer, and incubation, so “units” mean nothing until the assay is pinned down⁴.

4. Provenance. Keep the original value, original unit, lab, method, LOD/LOQ, and the transform applied. Every derived value should be reconstructible.

In your own data

Blood panel CSVs are the obvious surface. Check the unit column against the value range before trusting either: a creatinine of 0.9 is mg/dL, a creatinine of 80 is µmol/L, and a file that says µmol/L next to 0.9 is mislabeled. Values below LOQ often arrive as <0.5 or as a bare 0.5, and if your parser coerces to float you have just invented data. Keep censored values as a flag.

Proteomics is where the trap is subtle. Olink reports NPX, a log2 relative scale, so a “1.0 increase” is a doubling, and NPX is not comparable across panel versions or batches without bridging samples. SomaScan reports RFU. Neither is pg/mL. Do not mix them into one table without an explicit cross-platform mapping, and do not average them.

RNA-seq: decide once whether your unit is raw counts, TPM, or normalized counts. Differential expression tools expect raw counts. Cross-sample comparison of a single gene wants TPM or a size-factor-normalized value. Mixing the two across time points produces trends driven by library size.

Genotype data has its own harmonization problem: strand and allele coding. Arrays and sequencing call the same SNP on opposite strands, and A/T and C/G sites cannot be resolved by complementing alone, so you need allele-frequency matching against a reference panel. Genotype Harmonizer does this and handles format conversion⁵. Getting it wrong flips effect directions in a polygenic score silently.

Common mistakes we see: converting twice because two stages each applied a factor, harmonizing to a unit but not to a reference interval (units and intervals must travel together), and dropping method metadata after conversion so you can never audit a suspicious trend. Cross-archive projects hit the same wall at larger scale, and the durable fix is a shared ontology plus preserved source values rather than a one-off script⁶⁷.

Limitations

Harmonization cannot fix an assay that was not measuring the analyte. It cannot recover precision lost to a wide LOQ. Regression-based alignment is valid only over the concentration range you fit and degrades at the extremes, which is exactly where you care. Flow cytometry phenotyping shows how far gating and panel choices can move a result even with matched reagents⁸. And a harmonized number is still a number: interpretation against reference intervals, and any decision that follows, belongs with a clinician, especially for pharmacogenomic results where regulatory guidance is jurisdiction-specific⁹.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.