More biological data did not improve AI predictions of depression relapse

Adding brain scans, genetics, blood markers and gut microbiome data failed to improve prediction of depressive relapse in a study published on 21 October in JAMA Psychiatry. Its best model used a few clinical risk factors and reached 64.43% mean balanced accuracy, with a standard deviation of 6.63 percentage points.[1]

Lea Teutenberg, Nils R. Winter and colleagues studied 564 people with current or previous major depressive disorder in Marburg and Münster. Over roughly two years, 270 relapsed and 294 did not.[1]

The nine baseline data types also included cognitive tests and four structural or functional MRI modalities. Each added information measured at the outset; combining them did not improve performance.[1]

What the percentage means

Balanced accuracy averages two success rates: the share of relapses correctly identified and the share of non-relapses correctly identified. A model that predicted no relapse for everyone would score 50%, even if that happened to be the larger group. The reported figure describes how well a classifier separates the groups; it is not a forecast that an individual has that probability of becoming ill again.[2]

The clinical question matters because relapse prevention involves decisions about continuing care after symptoms improve. NICE’s existing guidance identifies recurrent episodes, residual symptoms, severe previous depression and continuing social stresses among factors that can increase risk. It recommends discussing further treatment with patients in light of their needs and preferences. A risk factor can inform that discussion without reliably identifying which particular person will relapse.[3]

The authors propose testing models that follow changing biological and clinical markers over time.[1]

Report an error or suggest a correction