Beyond BMI: Who Actually Needs Treatment

Every insurer, every prescriber and every telehealth intake in this market decides who gets treatment using BMI. A study in Nature Medicine built a model from thousands of candidate features and showed what that threshold hides: among people all above a BMI of 27, ten-year cardiovascular mortality ranged from 5.7% to 0.1%.^[1] A fifty-seven-fold difference, invisible to the number being used.

What the model does

Starting from thousands of tested features, the researchers identified the twenty most informative for predicting the future onset of eighteen obesity complications, and combined them into a single score. It stratified people into risk groups that behaved very differently over ten years, and it generalized to independent populations of both European and non-European ancestry.^[1]

Everyone in that table meets the same BMI criterion. One group faces more than fifty times the risk of another, and the criterion cannot tell them apart.

The finding that should change policy

The researchers then did something more useful than building another risk score. They applied it to participants in the SURMOUNT-1 tirzepatide trial and asked whether baseline risk predicted how well the drug worked.

Weight loss was similar across baseline risk groups.

That is the sentence that matters.^[1] If the highest-risk people lose about as much weight as the lowest-risk people, then treating by risk rather than by BMI costs nothing in effectiveness and gains everything in benefit. The same drug, given to someone facing 5.7% cardiovascular mortality instead of someone facing 0.1%, does far more good for the same money.

Predicted risks also decreased following treatment with tirzepatide — though that is a modeled decrease rather than observed events, which is a real distinction and worth keeping.

What the study does not establish

• It is a prediction model, not a treatment trial. It identifies who is at risk. It does not demonstrate that allocating treatment by this score improves outcomes, which would require its own trial.
• The risk reduction is modeled. “Predicted risks decreased” means the score fell, not that events were counted.
• Twenty features is still a lot to collect. A score that outperforms BMI only when you have twenty inputs is harder to use than a number from a scale and a tape measure.
• It does not tell an individual what to do. A risk group is a population, and our note on averages applies here as everywhere.

Why BMI persisted anyway

It is worth being fair to the thing being criticized. BMI is free, instant, requires no laboratory, and is calculable by anyone with a scale. Its universality is exactly why it became the gate.

But its weaknesses are well known — it does not distinguish fat from muscle, distributes differently across populations, and says nothing about where fat sits or what it is doing metabolically. What is new here is a demonstration of how much risk information a BMI threshold discards, and evidence that the discarded information would not cost you anything in treatment response to act on.