Can brain scans divide depression into distinct biological types?
A rigorous re-analysis shows that previously proposed biological subtypes of depression may be statistical artifacts of overfitting rather than true, distinct categories.
Source
Evaluating the evidence for biotypes of depression: Methodological replication and extension of
What they did
Researchers attempted to replicate a prominent study by applying its exact pipeline to a dataset of 187 patients with depression or anxiety. They selected 150 brain connectivity features that correlated best with 17 clinical symptoms, analyzed them using canonical correlation analysis, and performed clustering. They also introduced rigorous permutation tests and cross-validation to see if the resulting brain-behavior relationships and clinical subtypes were statistically reliable.
What they found
While initial analysis yielded seemingly strong brain-symptom correlations, permutation tests revealed these associations were not statistically significant and virtually disappeared in cross-validation. Additionally, the analysis identified a 3 cluster solution with a Calinski-Harabasz score of 109, but simulations demonstrated this clustering was no more distinct than what would occur by chance from a single continuous distribution. This suggests that depression is better represented as a continuous spectrum of symptoms rather than separate biological subtypes.
The limits
What it doesn't show
This study does not prove that biological subtypes of depression do not exist; rather, it shows that current methods are too weak to prove they do. Because the replication sample included a broader clinical group than the original study, some differences could stem from sample variation. Additionally, the study cannot rule out whether alternative imaging methods or different clinical measures might reveal genuine biological groupings.
Key terms
- Canonical Correlation Analysis (CCA)
- A multivariate statistical method used to find relationships and maximize correlation between two sets of multidimensional variables.
- Overfitting
- A modeling error that occurs when a statistical model matches its training data too closely, making it fail to predict new, unseen data accurately.
- Permutation Test
- A statistical method that shuffles the data to calculate a null distribution, allowing researchers to determine if an observed effect is statistically significant.
- Hierarchical Clustering
- A method of cluster analysis that seeks to build a hierarchy of groups based on how similar data points are to one another.
- Calinski-Harabasz (CH) Index
- A metric used to evaluate the quality of a clustering solution by comparing the variance between clusters to the variance within clusters.
- Functional Connectivity
- A measure of the statistical correlation or coordination of activity between spatially separate brain regions over time.
Flashcards
Want these cards to stick?
Save the deck to NoteFren and study it with spaced repetition.
Quiz yourself
What is the main objective of the Research Domain Criteria (RDoC) initiative in psychiatry?
Common questions
Why did the original study find distinct biotypes of depression while this study did not?
The original study did not use permutation tests or cross-validation to account for their initial step of selecting brain features, which heavily inflated their statistical significance and led to overfitting. When these rigorous checks are added, the seemingly distinct groups disappear.
Does this study prove that depression cannot be split into biological subtypes?
No, it does not prove subtypes do not exist; it simply demonstrates that the specific methods and dataset used here do not provide reliable evidence for them, suggesting depression looks more like a continuous spectrum in this sample.
What is the main clinical implication of these findings?
Instead of forcing patients into arbitrary biotype categories to predict treatment response, clinicians and researchers should use continuous, individualized biological and clinical scores, which preserve more information.
How did the researchers test whether the clusters they found were real or just random noise?
They simulated a single, continuous, normally distributed dataset with no true clusters and ran the clustering algorithm on it, showing that the random data produced cluster scores just as strong as those found in the real patient data.
More on Psychiatric neuroimaging