Fasting before a blood test: four results move, most do not

Someone has a blood draw booked for half past seven in the morning. At eleven the night before they are standing in the kitchen, hungry, holding a glass of orange juice and trying to remember what the clinic said. Eight hours. Or was it twelve. The appointment card says fasting and nothing else, the internet says eight to twelve hours with great confidence and no explanation, and the actual question — whether a sip of juice at eleven ruins the test, or only moves the one measurement nobody is looking at — has no answer anywhere they can find.

That is the question a new analysis in JAMA Internal Medicine speaks to directly, and its answer is narrower and more useful than the headline it produced. The paper, published on 1 November 2026 by Sunghwan Shin and colleagues, did not ask whether fasting matters. It asked how much the length of a fast changes each individual test, and then measured those changes against the standard laboratories already use to decide whether a difference is large enough to count as a difference at all.

The team took the laboratory information system of a 670-bed secondary-care university hospital in the Seoul metropolitan area and pulled adult outpatient results from February 2021 to December 2023 where the time since the last meal had been documented. All of the blood was drawn at the hospital's central phlebotomy unit. The final dataset comprised 9,755,547 test results from 101,148 patients across 372 test items, analysed between January 2024 and July 2026. It is a retrospective, single-centre, cross-sectional study: nobody was assigned to fast for a particular length of time, and nobody was followed forward.

Fasting duration — the interval between the last meal and the draw — was sorted into three bands: under eight hours, eight to twelve hours, and more than twelve. Group means were adjusted with linear mixed-effects models carrying a random intercept for each patient, accounting for age, sex and the department that ordered the test. Because the same patients often came back, the authors also ran paired within-person comparisons, which hold the individual constant and ask whether one person's own result moved when their fasting time did.

The last decision is the one that makes the paper readable. Rather than reporting which differences were statistically significant — which, at nearly ten million results, almost everything would be — the authors compared each difference against analytical performance specifications derived from the European Federation of Clinical Chemistry and Laboratory Medicine's biological-variation data. That is the laboratory field's own total allowable error: the size of a difference below which a laboratory treats a result as the same result. Within the under-eight-hour band they also tracked results at thirty-minute intervals, to see how measurements drifted in the hours after a meal.

Comparing the three bands against one another, 27 analytes showed differences exceeding the allowable-error standard. Only 10 of those survived the paired within-person analysis, which the authors read as a sign of confounding — in other words, much of the apparent effect came from the fact that people who fast for different lengths of time are different people, have different tests ordered, and arrive from different departments.

The analytes the paper names are glucose, triglycerides, gamma-glutamyl transferase and lipase, each running higher when the last meal was less than eight hours before the draw: glucose by 10.8%, triglycerides by 20.1%, gamma-glutamyl transferase by 14.8% and lipase by 36.4%. The abstract reports those four figures without stating what they are percentages of, so the direction and the ordering are solid while the precise magnitudes are not numbers to carry away as exact.

The finding most likely to change what happens at a phlebotomy desk is the comparison between the two longer bands. No analyte exceeded the allowable-error standard between the eight-to-twelve-hour group and the group fasting more than twelve hours, and the authors state in their conclusion that no estimable within-person difference between those two groups exceeded an available specification either. Inside the under-eight-hour band, glucose and lactate fell significantly as the fast lengthened, while lipid concentrations showed no significant linear trend over the same hours.

The authors' own conclusion is the sentence to keep: these findings support analyte-specific rather than uniform fasting requirements. The instruction is doing real work for a short list of tests and almost none for the several hundred others.

A study this size can measure almost anything, which is precisely why its honesty rests on the two checks the authors built in. The first is the allowable-error comparison, which asks not whether a difference exists but whether it is large enough for a laboratory to call it a different result. The second is the within-person analysis, and it is the one that took 27 flagged analytes down to 10.

That fall is itself a measurement of how misleading record-based comparisons of fasting can be. Patients who arrive having eaten are not a random sample of patients: they are likelier to be there for a test nobody asked them to fast for, likelier to be slotted between other appointments, and likelier to belong to a department that orders a different panel. Holding the person constant strips much of that away, and most of the effect goes with it. An analysis that reported only the between-group half — the half most easily turned into a headline — would be reporting a number these authors could not confirm.

What that check cannot do is reach outside the hospital it was run in. Instruments, reference methods and local ordering practice all shape where an analyte lands, and a single centre means a single set of those. The abstract is also silent on how fasting duration came to be documented, which is the one input the entire analysis rests on.

The practical weight of this paper is not in the four analytes that moved. It is in the 372 test items it looked at and the handful that turned out to care. A uniform fasting instruction is cheap to print and expensive to follow: it produces missed appointments, repeated draws, early-morning travel on an empty stomach for people who are elderly, pregnant, diabetic or simply working, and a quiet population of patients who ate something and said nothing about it. If the requirement is genuinely analyte-specific, a good deal of that cost buys nothing.

It also reframes the eleven-o'clock question in the kitchen. What matters is not a count of hours but which measurement is sensitive to a meal, and a glucose test, a lipid panel and a liver enzyme are not in the same position as the great majority of what a routine draw measures. That is a property of the requisition rather than of the patient's discipline the night before. Clinicians weighing a non-fasting sample consider which analytes were ordered, whether a result sits near a threshold they are reading against, and whether a repeat draw costs the patient more than the uncertainty it would remove — questions this study informs without settling.

This is a paper that earns its conclusion by refusing the easy version of its own finding. The temptation at nearly ten million results is to report everything that reached statistical significance; instead the authors measured each difference against the laboratory field's own allowable-error standard and then, having flagged 27 analytes, told their readers that only 10 held up when the same person was compared with themselves. The claim worth carrying from this work is the narrow one — that between a fast of eight to twelve hours and a fast of more than twelve, nothing the authors could estimate exceeded an available specification. The broader call for analyte-specific rather than uniform fasting requirements is their argument rather than their measurement, however reasonable it looks from the data. And the study cannot license a general claim that fasting does not matter: four analytes moved, one of them by a third, and the full list of ten was never published. What it does is shift the burden onto a uniform instruction that has never been justified test by test — one retrospective hospital dataset will not settle a practice this entrenched, and it should not have to carry that weight alone.