Abstract

Assessing the Generalizability and Stability of Biologically Based Subtypes of Depression

Dunlop K, Grosenick L, Downar J, Vila-Rodriguez F, Gunning F, Daskalakis Z, Blumberger DM, Liston C
61st Annual Meeting of the American College of Neuropsychopharmacology. 2022.

Background: Major depressive disorder (MDD) is associated with considerable symptom variability; a comprehensive understanding of this variability may lead to individualized intervention approaches and therefore improved treatment response rates. Recent efforts by our group and others have sought to understand symptom heterogeneity in MDD using functional neuroimaging. Although these recent studies mark significant progress in understanding heterogeneity in MDD, replication and validation of these results is critical given the limitations imposed by sample size and depth of clinical characterization. To address these limitations, this study has three aims. First, we sought to understand the neurobiological basis of MDD symptom heterogeneity by extending our earlier work defining robust and reproducible brain-behavior dimensions to new data. An L2-norm regularized multivariate model was generated using a large MDD dataset recruited from a single site and incorporated additional items assessing anhedonia and anxiety symptoms. Second, we tested for the existence of MDD subtypes and evaluated their stability and reproducibility. Third, we characterized these MDD subtypes regarding atypical resting-state functional connectivity (RSFC), clinical symptoms, and antidepressant response to non-invasive brain stimulation.

Methods: L2-regularized canonical correlation analysis (RCCA) was evaluated in a large, single-site MDD dataset (n = 328, 215 female (65.6%); mean age=40.35 ± 12.05 SD) using RSFC and clinical symptomatology. First, to optimize three RCCA hyperparameters, we performed a nested grid search (with training, validation, and test splits); the optimal hyperparameter combination was defined as the highest median canonical correlation in held-out validation data for the first dimension. Next, we examined the stability and hold-out performance of dimensions (on held-out test data not used for training or validation) and tested for significant dimensions using random permutation testing. Afterwards, we generated a final optimized RCCA model and evaluated the performance and stability of hierarchical clustering. Upon identifying the optimal clustering solution, we characterized latent variables representing co-occurring RSFC and symptomatology, and symptom/RSFC differences by subtype. Lastly, we identified subtype differences in repetitive transcranial magnetic stimulation response and remission rates.

Results: The performance and stability of the first three RCCA dimensions were significant (p < 0.05, random permutation test). These three dimensions represented: depressed mood, and thalamic and default mode RSFC; anhedonia, and cingulo-opercular and higher-order visual and network RSFC; and insomnia, and sensorimotor and posterior insula RSFC, among other connectivity features. Hierarchical clustering identified four significant depression subtypes (p < 0.05, random permutation test), each with distinct clinical symptom profiles, abnormal RSFC patterns, and responsivity to repetitive transcranial magnetic stimulation (rTMS) over the dorsomedial or dorsolateral prefrontal cortex. Subtypes with lower anhedonic symptoms were most responsive to rTMS. Subtypes did not differ by age or sex.

Conclusions: In an extension of our previous work, we sought to characterize regularized CCA and clustering performance in a large, single-site MDD dataset. RCCA yielded three significant, stable and generalizable brain-behavior dimensions that resembled well-documented MDD symptom-brain associations, and four categorical subtypes. Both categorical and dimensional approaches to parsing heterogeneity may be beneficial in different contexts. We note several study design choices that may affect RCCA models, including participant inclusion/exclusion criteria, medication use, and choice of symptom severity measures. Taken together, these results represent an important step forward in assessing data-driven subtyping methods and provide evidence that RCCA is an effective tool to identify stable and generalizable associations between RSFC and behavior.