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

A German cohort of 564 adults with current or past major depression was assessed at baseline across nine data types: clinical and psychometric scores, four MRI modalities, genetics, cognitive testing, blood markers and the gut microbiome. Machine learning models then tried to predict who would relapse over about two years; 270 did.

The best model reached a balanced accuracy of about 64%, where 50% is chance. That best model used only a handful of clinical relapse risk factors. Adding imaging, genetic, blood or microbiome data did not improve prediction.

This is an honest negative. It suggests that, measured once, biological markers do not yet tell a clinician more than a careful history about who will relapse. Accuracy in the mid-60s is too low to guide decisions for an individual patient.

For practice, the message is reassuring rather than new: the clinical picture already captures most of what is predictable, and there is no test to order.

Do not order imaging, genetic or microbiome tests to predict depressive relapse.
Record the clinical relapse markers: number of past episodes, residual symptoms and age at onset.
Use the clinical history to decide on maintenance treatment and follow-up intensity.
Treat commercial relapse-prediction tests with caution until they beat clinical factors.

Why it matters

It suggests the expensive biomarker route to personalised relapse prevention has not yet earned its place in clinic.

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 and 64% is modest: about a third of predictions would still be wrong. The models were built and tested in one cohort, so real-world accuracy may be lower.