Lukasz Regulski

Research · Open data

Personal blood norms

A normal range built from a person’s own past results beat the population range for 47 of 51 routine blood tests.

Your normal, not everyone's Left, an illustration: one person's past blood results sit in a narrow personal range inside the wide population range, and a new result falls outside the personal range while still inside the population one. Right, the result: for 47 of 51 blood tests the personal range scored better than an age- and sex-matched population range; ratios run from 0.37 to 1.03, where below 1 favours the personal range. population range your range past results new Your normal, not everyone’s 0.4 0.6 0.8 1.0 lower = personal better Tighter in 47 of 51 tests

Every blood result comes with a reference range: the values most healthy people fall within. But people differ from each other far more than they differ from themselves. Your creatinine may have sat near the bottom of the range for twenty years. A value near the top is normal for the population and a large change for you.

So the question: if you build a “normal” from a person’s own previous results, is it better than the population range?

I used MIMIC-IV, a research database of de-identified records from a Boston hospital: 195,000 people and 29 million outpatient test-days across 51 routine tests (blood count, chemistry, lipids), taken away from any hospital stay. For each result I predicted a 95% range from that person’s earlier results and compared it with a range for their age and sex. The score rewards narrow ranges and penalises misses, so neither side wins by being vague.

What it showed

  • The personal range won for 47 of 51 tests, from the first earlier result. For creatinine (a kidney marker) the personal score was less than half the population’s (0.44 of it); HDL cholesterol 0.58, haemoglobin 0.82. The four exceptions are the rare white cells (eosinophils and basophils), whose counts are mostly zero.
  • The trick is shrinkage. The personal range leans on the population when a person has few results and on the person as results pile up. A person’s raw average and spread, without that, did worse than the population range for 44 of 51 tests.
  • Long histories help less than expected. People drift over years, and a model that assumes a fixed set point starts to miss: coverage fell from 96% at the first earlier result to 91% by the twentieth, while the population range stayed near 94–96%. The next model tracks drift.
  • Unusual for you is not the same as dangerous. Distance from your own range did not predict death within a year better than distance from the population range, and for red-cell tests it did worse. A chronically anaemic person’s own normal is anaemia, and the anaemia is what matters.

Why it is interesting

It is the frailty idea again: judge a measurement against the person’s own history. As a description of a person, it clearly works. Whether it improves decisions is still open, and the answer differs by test. These are hospital patients, so the next step is data from people who mostly stay well.

The picture on the left of the image is an illustration, not real data.

How it was done

Data. MIMIC-IV laboratory results: outpatient draws at least 30 days from any hospital stay, one value per person, test and day. The 51 tests are the most common routine blood tests, leaving out blood gases and tests mostly drawn in hospital.

The task. For every result where the person had an earlier one, predict a 95% range from the earlier results only. Models were fitted on four fifths of the people and tested on the rest, five times over, so no one was predicted by a model that had seen them.

What competed. The lab’s printed range; a population range for the person’s age and sex; and personal ranges built from the person’s history, with and without shrinkage, as a trend line, and by gradient boosting.

Shrinkage. The personal range leans on the population when a person has one or two results and moves towards the person as results accumulate. This is what makes it usable from the first repeat; without it, personal ranges lost to the population range for most tests.

Scoring. The interval score, which adds the width of a range to a heavy penalty for every result that falls outside it, so a range cannot win by being either vague or overconfident.

Also worth knowing. A model that also sees the other tests drawn that day predicts even better, because many blood count values determine each other, but that answers a different question from “what is normal for you”. The gain was largest for tests where people differ most from each other and least from themselves.

Limits. MIMIC cannot separate a person’s own variation from lab measurement error, and the outcome test looked at the size of a deviation, not its direction. A drift-aware model and a signed analysis come next.

What happens as a person's history grows Top: median coverage across 51 tests of the nominal 95 percent range. The personal range starts at 96.4 percent with one earlier result and falls to 91.3 percent by the twentieth; the population range goes from 96.1 to 94.3 percent. Bottom: ratio of interval scores, personal over population. Creatinine is lowest, about 0.31 at three results and 0.45 by twenty; haemoglobin stays near 0.77 to 0.84; HbA1c rises from 0.66 to about 0.9 to 1.0 after fifteen results. Share of next results inside the 95% range 90% 92% 94% 96% target 95% population personal Personal score ÷ population score (lower = personal better) 0.2 0.4 0.6 0.8 1.0 no gain Creatinine Haemoglobin HbA1c 1 5 10 15 20 earlier results the person already had
Top: median across 51 tests of the share of next results that fell inside each nominal 95% range. Bottom: mean interval score of the personal range divided by that of the age/sex population range, for three tests. Each point pools every person with that many earlier results: from about 100,000 people at one earlier result to 11,000–12,000 (creatinine, haemoglobin) and about 1,000 (HbA1c) at twenty, which is why HbA1c is noisier. MIMIC-IV summary statistics.