Thread by @NilsRWinter

Thrilled to see this one published in @JAMApsych. Here's a short thread discussing our findings on neurobiological #biomarkers of #depression. jamanetwork.com/journals/jamap…

In a large, bi-centric study (#FOR2107) on depression (n=1809), we quantified the univariate neurobiological differences between healthy subjects and patients with major depression across a number of MRI modalities (#sMRI, task and rest #fMRI, #DTI).

We show that the univariate difference between the two groups is remarkably small for all modalities, all explaining less than 2% of variance.

Within a precision psychiatry framework, individual predictions become increasingly important. We therefore estimated the upper bound for the single subject classification accuracy of the variables showing the largest group difference (per modality).

Univariate classification peaks between 54 to 56% (remember that 50% is random guessing). Looking at the distributional overlap in these variables showing the strongest effect, the similarity between the two groups becomes evident immediately (overlap between 87% and 95%).

This pattern of results does not change when considering patients with acute or recurrent symptomatology only. We therefore conclude that the potential of univariate measures to function as MDD biomarkers is surprisingly small.

Although there is a tremendous amount of suffering in patients with MDD and a clear clinical/behavioral difference to healthy controls, why don't we see stronger effects and higher potential for single-subject predictions in univariate neurobiological measures?

I strongly believe we must try to understand and resolve this apparent discrepancy if we want to have a chance of changing the clinical reality of patients. E.g., many argue that MDD is not a heterogeneous disorder, both clinically and neurobiologically.

Especially in a case-control design, we might be limited to explaining only the variance that is shared across patients (and due to the heterogeneity, this might not be a lot). In this case, normative modeling could help overcoming some of these limitations.

Finding more (clinically, biologically) homogeneous subtypes might help as well, but this is far from trivial, yet worth trying. Also focusing on single symptoms to reduce clinical heterogeneity could improve results. Especially symptom networks are a promising way forward.

Lastly, we only report univariate (i.e., based on a single measure) differences and classification accuracy. Certainly, machine learning folks will agree that these models underfit massively and that we need to look at multivariate biomarkers. We agree and will do so now.

And finally, big thanks to all the great people from #FOR2107 #TraPLab_Münster in Münster and Marburg who collected this massive dataset and to @TheRealTimHahn1 and Udo Dannlowski for supervising this project.