Are biological age tests actually accurate?

Brain-scan-based models for detecting major depression perform no better than chance. Researchers trained no fewer than 4 million machine-learning models on three types of brain scans: structural MRI, functional MRI, and diffusion tensor imaging. Average accuracy ranged from 48% to 62%, while flipping a coin already gives you 50%. The distributional overlap between people with depression and healthy people was 87 to 95%, meaning the brain scans of the two groups are barely distinguishable from one another. At the individual level, such a test therefore provides no usable diagnosis.

This large-scale modelling study illustrates a broader principle: high technological complexity, involving millions of models and expensive scans, does not guarantee clinical usefulness. The fact that something sounds impressive says nothing about whether it works for you as an individual. It is the difference between what a method does in a group and what it tells us about one specific person.

A blood test measuring the protein phospho-tau217 as an indicator of Alzheimer's disease presents a very different picture. This test has been validated in multiple patient populations, both in general practice and in specialised centres, and shows considerably higher diagnostic accuracy than the brain scans for depression. Phospho-tau217 is not a brain-age test in the strict sense, but it does fall within the broader category of objective biological measurements for brain disease.

The distinction that matters for the reader is therefore not whether a test is called a 'biological age' test, but whether individual-level accuracy has been demonstrated in multiple independent settings. For depression via brain scans, that evidence is clearly negative. For Alzheimer's via blood phospho-tau217, the results are validated and promising.