Combining brain scans, genetics, blood markers, gut microbiome data and cognitive tests did nothing to improve the prediction of who would relapse into depression over two years compared with a handful of clinical risk factors, according to a German study of 564 patients published October 21, 2026, in JAMA Psychiatry. Even the best model was right only about two-thirds of the time.
The peer-reviewed study, with first authors Lea Teutenberg, Nils R. Winter and Paula Usemann and senior author Tilo Kircher of Philipps-Universität Marburg, together with colleagues at the University of Münster and other institutions, used data from the Marburg-Münster Affective Disorders Cohort Study. Participants with current or lifetime major depressive disorder and no history of neurological or severe physical illness were recruited from psychiatric inpatient and outpatient departments and from the general population at two sites. Baseline assessments took place between October 2014 and May 2019, with follow-up about two years later; the machine learning analysis was conducted between July 2024 and June 2026.
Of the 564 patients (65.1% women; mean age 37.3 years), 270 relapsed during follow-up and 294 did not. The researchers trained models on nine types of data: clinical and psychometric measures, four structural and functional MRI modalities, genetics, neurocognition, blood markers and gut microbiome profiles. The primary performance measure was balanced accuracy, which averages the rate of correctly identified relapses and correctly identified non-relapses, so that 50% equals chance.
The study was naturalistic, meaning participants were followed under whatever care they received rather than being assigned to a treatment protocol, and the outcome was a simple binary classification: relapse or no relapse within the two-year window.
Mean balanced accuracy across models ranged from 47.2% to 64.4%. The best-performing model relied solely on a few established clinical relapse risk factors; adding the other modalities did not improve performance. The authors conclude that the models reached only moderate accuracy, with limited scope for guiding decisions about individual patients, and that prediction was driven largely by known clinical factors. Funding and conflict-of-interest statements are not shown on the journal’s abstract page; the abstract is also indexed on PubMed.
Why it matters
A reliable way to identify which patients are likely to relapse would help clinicians target maintenance treatment and follow-up, and much of the investment in precision psychiatry has gone into biological measurements on the premise that neuroimaging, genomics and other markers capture signals that a clinical interview cannot. This study, with a large sample, naturalistic follow-up and an unusually broad set of data types, is a direct test of that premise in depression, and the answer was negative. The journal published it in its “AI in Psychiatry” section.
The result is a caution for digital and precision psychiatry more generally: expensive measurements are not automatically informative, and a prediction model built from biomarkers should be judged against the simplest clinical alternative before it is used. For developers of prediction tools, it is also a reminder that a headline performance figure needs context: a balanced accuracy of 64% is 14 points above the 50% that a model with no information would achieve. The authors suggest that a single baseline snapshot may be the wrong approach, and that dynamic modeling, which tracks how multimodal markers change over time alongside outcomes, may do better.
The caveats
Balanced accuracy in the mid-60s means the best model misclassified roughly one in three patients, which is far from clinically useful, but the study was conducted at two sites in one country and relapse was predicted from a single time point. The abstract does not give the operational definition of relapse or the specific clinical factors used in the best model, and no external validation in an independent cohort is reported on the abstract page. The cohort excluded people with neurological or severe somatic conditions, who may differ from the broader population of people with depression. The full article may contain detail, including funding, that is not available on the summary page.
This article is for general information and does not replace advice from a qualified clinician. Corrections: email the editor using the address on the Contact page.
