Design: Naturalistic longitudinal prognostic study with machine learning
Population: 564 adults with current or lifetime major depression at two German sites
Primary outcome: Balanced accuracy for predicting relapse within about 2 years
Effect: 47.2% to 64.4%; best model used clinical risk factors alone

A prognostic study in JAMA Psychiatry followed 564 people with current or lifetime major depression from a German cohort for about two years; 270 relapsed. Machine-learning models drew on nine kinds of baseline data, including clinical and psychometric measures, four MRI modalities, genetics, neurocognition, blood markers and gut microbiome.

Balanced accuracy for predicting relapse ranged from 47% to 64%. The best model used only a few clinical relapse risk factors, and combining modalities did not improve it.

The practical message is reassuring for the clinic: the information that best predicts relapse is already in the history. The authors judge even the best model too weak to guide decisions for an individual patient.

Take a careful episode history — number of prior episodes and residual symptoms remain the most useful relapse markers.
Do not order imaging, genetic or biomarker tests to predict relapse in depression; they added nothing here.
Base decisions on continuing antidepressants on clinical risk factors and patient preference.
Tell patients asking about relapse tests that no current test predicts it reliably for an individual.

Why it matters

This challenges the expectation that richer biological data will soon personalise relapse prevention.

The statistics, in plain English

Balanced accuracy averages how well a model identifies those who relapse and those who do not; 50% is a coin toss. At best 64%, the models were modestly better than chance — useful for research, not enough to decide treatment for one person.