Normative modelling · Lobeworks/17

Normative modelling is a method of placing one person's brain measurement on the distribution of a large reference population, as a growth chart places a child's height, so that a patient is described by how far and where they depart from the expected range instead of by the average of the diagnostic group they belong

Normative modelling. Normative modelling is a method of placing one person's brain measurement on the distribution of a large reference population, as a growth chart places a child's height, so that a patient is described by how far and where they depart from the expected range instead of by the average of the diagnostic group they belong to.

It answers an old problem. Since Emil Kraepelin split dementia praecox (today's schizophrenia) from manic-depressive insanity in 1899, by course and symptoms, psychiatry has looked for brain markers of its categories, and the standard tool has been the case-control study: scan patients and controls, compare the means. The design assumes each diagnosis is one biology, and the means it finds are small and hide the spread. A normative model instead fits, on thousands of healthy scans, the expected value and spread of each measure (cortical thickness, regional volume) as a function of age, sex and scanner, then scores each individual as a centile or deviation. Applied across six diagnoses, the deviations of individual patients overlapped poorly: the same region was affected in fewer than 7 % of people sharing a diagnosis, while the deviations converged more on shared circuits and networks.

Brain charts are its large-scale form. Bethlehem and colleagues pooled 123,984 MRI scans from 101,457 people in more than 100 studies, covering ages from 115 days after conception to 100 years, and drew lifespan curves of grey and white matter volume as paediatric growth charts do for height and weight.

It is the brain version of anomaly detection against a reference distribution: a measure is read as a percentile of its peers, exactly as a maintenance team flags a pump whose vibration sits at the 99th percentile of pumps of its age and duty, rather than comparing it with the average failed pump. The normal distribution is the simplest reference; real charts fit skewed, age-dependent curves.

The reference decides what counts as normal. About 1 % of the brain-chart scans came from South America and Africa together, and scanner and site effects must be modelled out, so a deviation can reflect the sample as well as the person.

A deviation is a description, not a diagnosis: it says a brain is unusual for its age and sex, and leaves open whether that matters clinically. Its value is phenotyping, grouping people by their biology across diagnostic labels.

Compare each brain with the distribution, not with the average patient.

Group means can differ while almost no individual shares the difference; centiles keep the individual in view.

Questions: Why has no single brain marker of depression been found? Because the diagnosis gathers different conditions under one name: five of nine symptoms allow at least 227 combinations, and the routes in (stress, inflammation, genetics, illness) differ between people. Comparing patients with controls finds real but small group differences, such as a hippocampus about 1.2 % smaller in recurrent illness, by which no single patient can be diagnosed. Normative modelling changes the question: instead of looking for one difference shared by the group, it measures how each individual departs from a reference population, and finds that patients scatter.