Comparing data-driven subtypes of depression informed by symptom and neuroimaging data

469 

Abstract Submission 

Kayla Hannon¹, Luca Balogh², Fyzeen Ahmad³, Petra Lenzini¹, Aristeidis Sotiras¹, Janine Bijsterbosch¹

¹Washington University in St Louis, St Louis, MO, ²University of Amsterdam, Amsterdam, North Holland, ³University of Minnesota, Minneapolis, MN

Kayla Hannon  
Washington University in St Louis
St Louis, MO

Luca Balogh  
University of Amsterdam
Amsterdam, North Holland

Fyzeen Ahmad  
University of Minnesota
Minneapolis, MN

Petra Lenzini  
Washington University in St Louis
St Louis, MO

Aristeidis Sotiras  
Washington University in St Louis
St Louis, MO

Janine Bijsterbosch  
Washington University in St Louis
St Louis, MO

The wide heterogeneity of depression (both clinically and neurobiologically) points to the presence of subtypes within the disorder. However, efforts to subtype depression have failed to converge on a consensus. Our study aims to compare several previously developed data-driven depression subtyping approaches informed by either symptom or neuroimaging data within the same subject space. We leverage the rich symptom and neuroimaging data in the UK Biobank (UKB). We evaluate similarities in resulting subtypes on subject cluster assignment agreement and sensitivity to differentiation of clinical and biological phenotypes.

Subtyping approaches: We applied the subtyping approaches of two studies that clustered on symptom data [Maglanoc et al 2019 & Lamers et al 2010]. Briefly, Maglanoc et al performed Gaussian mixture discriminant analysis on symptom questions using Bayesian Information criterion to determine the optimal cluster solution. Lamers et al performed latent class analysis on symptom questions with Akaike Information criterion for cluster optimization. We also applied the subtyping approaches of two studies that clustered on functional neuroimaging data [Price et al 2017 & Drysdale et al 2017]. Price et al performed a group iterative multiple model estimation (GIMME) on functional networks, determining clusters using the walktrap algorithm. Drysdale et al performed hierarchical clustering on functional nodes related to depression measures, with an optimal solution based on the variance ratio criterion. Please see the pre-registration, accepted as a stage 1 registered report, for more details [Hannon et al 2023]. We applied these approaches in the same UKB sample (N=2299) who have been identified to have moderate to severe recurrent depression [Smith et al 2013]. We evaluated the agreement of the resulting cluster solutions between approaches using the Adjusted Rand Index (ARI).

Sensitivity to Phenotype Differentiation: To determine which domain of phenotypes each subtyping approach is sensitive to, we evaluated the cluster solutions on the same demographic information, clinical measures, and structural neuroimaging measures (only including measures not used to drive any of the clustering approaches). We performed ANOVAs within each approach for a within-approach evaluation of phenotype differentiation, using false discovery rate correction across all phenotypes. We performed linear models with regressors for each subtype to extract variance explained (R2) measures for each subtyping approach. The resulting R2 were used to compare subtyping approaches and assess phenotype sensitivity. We determined significance by creating a null distribution of permuted ∆R2, then taking the maximum permuted ∆R2 across all possible pairs of combinations for each of 2000 shuffles. A ∆R2 was significant accounting for multiple comparisons if larger than the equivalent maximum permuted ∆R2.

The optimal subtype solutions ranged from 2 to 8 clusters (Fig 1A). The agreement between clustering approaches was minimal (ARI<0.023; Fig. 1B), indicating that there was functionally no overlap of cluster solutions between any subtyping approach, even when clustered on similar data. All of the cluster approaches showed significant within-method cluster differences (Fig 2A&B). The two symptom-based cluster approaches were more sensitive to clinical and imaging phenotypes than the neuroimaging-based cluster approaches (Fig 2C&D), although the phenotype sensitivity patterns were largely inconsistent (Fig 2E).

We find a lack of agreement on depression subtypes across approaches, which demonstrates the impact of analytical decisions in subtyping efforts. Different subtyping approaches were sensitive to different phenotypes, indicating they were parsing different domains of heterogeneity. In all, this work indicates different subtyping approaches capture different sources of heterogeneity that do not have a direct relationship to other sources of heterogeneity.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Classification and Predictive Modeling ²

Anatomical MRI

BOLD fMRI

Diffusion MRI

FUNCTIONAL MRI

Pre-registration

Psychiatric Disorders

STRUCTURAL MRI

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC