Why a blank sample doesn't read zero

The three numbers are the limit of blank (LoB), the limit of detection (LoD) and the limit of quantitation (LoQ). In clinical chemistry the standard method for setting them comes from the Clinical and Laboratory Standards Institute guideline EP17. The clearest short explanation of it is a 2008 review in The Clinical Biochemist Reviews by David Armbruster and Terry Pry, both of Abbott Diagnostics, and that review is the main source for what follows.

Nothing still makes a signal

A blank is a sample known to contain none of the thing being measured. Detectors have noise, reagents have background, and the matrix can contribute a little signal of its own, so replicate blanks scatter around some value above zero. The review puts it plainly: a blank sample "can produce an analytical signal that might otherwise be consistent with a low concentration of analyte."

EP17 defines the limit of blank as the highest apparent concentration you expect to see when testing replicates of a blank. The working formula is:

LoB = mean of the blanks + 1.645 × SD of the blanks

If the blank results follow a bell curve, 95% of them fall below that line. The other 5% land above it, which means a truly empty sample will occasionally look like it contains something. That is a false positive, and EP17 accepts it at a known, fixed rate rather than pretending it away.

The shortcut that measures nothing

The traditional way to set a detection limit was quicker. Run about 20 replicates of a blank or zero calibrator, take the mean, add two standard deviations (or three, four, even ten for a more conservative figure) and call that the LoD.

The problem, as Armbruster and Pry point out, is that this method never tests a sample that actually contains the analyte. There is no evidence that a low concentration produces a signal you can tell apart from the blank. They quote an older paper by Needleman and Romberg on drugs-of-abuse testing that sums it up in one line:

"It defines only the ability to measure nothing."

What EP17 does instead

EP17 builds the detection limit on top of the blank limit, using real low-concentration samples:

LoD = LoB + 1.645 × SD of a low-concentration sample

Set this way, a sample sitting at the LoD gives a result above the LoB 95% of the time. The other 5% read as if nothing were there, which is a false negative. So the LoD is defined by both error rates at once: about 5% of blanks look positive, and about 5% of samples at the LoD look negative.

It helps to see why the gap between the two limits is needed. Imagine a sample whose true concentration sits exactly at the LoB. If its results scatter symmetrically, roughly half will fall below the LoB and get called "not detected." That is my illustration, assuming a bell curve, not a figure from the review. It is why the LoD has to sit higher than the LoB, by enough of the low sample's own scatter to push misses down to about 1 in 20.

The review describes the testing load. A manufacturer establishing these limits is expected to use two or more instruments and reagent lots, with a recommended 60 replicates. A lab verifying the manufacturer's claim can use 20. The check is simple: if more than 5% of results from a sample at the claimed LoD fall below the LoB (roughly 1 in 20), the LoD is too low and has to be re-estimated with a higher concentration.

Detected is not the same as measured

Clearing the LoD means you can say something is there. It does not mean the number attached to it is any good. That is what the limit of quantitation is for: the lowest concentration where the result also meets predefined goals for bias and imprecision. It can equal the LoD, or sit well above it, but it can never be lower.

A related clinical term is "functional sensitivity," the concentration at which the assay's coefficient of variation reaches 20% (or another preset level). The review notes it was developed for TSH assays, where telling low-normal from suppressed values at very low concentrations matters for diagnosis.

One more term to watch: "analytical sensitivity" is sometimes used to mean the slope of the calibration curve, and sometimes used loosely as a synonym for LoD. The review advises against the second usage, because the LoD can sit below the range where the calibration curve is even valid.

Why the screen says zero

There is a practical wrinkle. Most clinical analyzers do not show the raw signal; they convert it to a concentration, and many report anything below a fixed cutoff as zero. The review notes this artificially inflates the number of "zero" results and recommends using the raw signal to establish the LoB. So a column of zeros from a blank run is not evidence of a quiet instrument. It may just be the software rounding the noise away.

Anyone who has worked with a spectrometer will recognize the general point: a blank is never perfectly flat, and the honest question is not whether it reads zero but how far it wanders.

Where it matters

For many routine tests none of this changes a clinical decision. The review uses glucose and cholesterol as examples: their decision levels sit so far above the assays' lower limits that the bottom end rarely comes into play. The limits matter for tests that ask "is it there at all," and the review names drugs, troponin and hCG, plus low-end hormone measurements like TSH where the LoQ governs. (Report 062 covers the same family of numbers in environmental testing, where the regulatory terms are method detection limit and reporting limit.)

How to read a detection claim

Was it built from blanks alone? A "mean plus 2 SD of the blank" figure says how noisy the blank is, not whether a low sample can be seen.

What are the error rates? Under EP17's parametric approach, about 5% false positives at the LoB and about 5% false negatives at the LoD. A result just above the LoD is a yes with a known chance of being wrong.

Is the number quantitative? Between the LoD and LoQ the analyte is present but the value may not meet accuracy goals.

How many instruments and lots? A limit set on one analyzer with one reagent lot will usually look better than the field performance it is supposed to predict.

The honest limits on this

The Armbruster and Pry review is short, written by industry scientists, and describes the first edition of EP17. The current edition, EP17-A2 (2012, reaffirmed 2017 and recognized by the FDA), is a paid document and I have not read its full text; the review itself says the guideline contains more statistical detail, including non-parametric methods for data that are not bell-shaped. The 1.645 multipliers above assume a normal distribution and should be read as the standard parametric case, not the only way the guideline allows.