Identifying stable cognitive clusters in a cohort of patients with major depressive disorder: A BrainDrugs-Depression study

Introduction

Major depressive disorder (MDD) is a heterogeneous disorder that can be difficult to treat, complicating clinical decision-making [1][2]. Data-driven approaches have sought to identify MDD subtypes to inform targeted treatment strategies; however, findings remain inconsistent. Cognitive deficits, increasingly recognized as central to functional recovery and treatment outcomes, may represent a critical but underexplored dimension for subtyping, as impairments can persist beyond core depressive symptoms and adversely affect daily functioning and response to treatment [3][4].

Aims

This study examines whether cognitive clusters can be identified in first-episode MDD [5] using both cold, hot, and social cognitive tasks within a multi-algorithm clustering framework. We further evaluate the robustness of the identified clusters using cross-validated consensus methods and assess whether those clusters differ in sociodemographic and clinical characteristics, as well as outcome after treatment-as-usual in out-patient treatment.

Methods

A total of 320 patients with first-episode MDD (70% female; mean age 30.4 ± 11.3 years) underwent pretreatment cognitive testing across cold domains (reaction time, psychomotor speed, executive function, working memory, and verbal memory), hot domains (biases in emotional recognition, misattribution, and detection thresholds), and social domains (guilt and shame rating). Five clustering algorithms (hierarchical clustering, spherical k-means, affinity propagation, HDBSCAN, and Gaussian-mixture models) were applied and evaluated using five-fold cross-validation after which a final stable consensus cluster solution was derived. Differences in sociodemographic and clinical characteristics between clusters (e.g., The Inventory of Depressive Symptomatology) were compared with ANOVA and χ² tests. Treatment outcome was %-reduction in depression severity at first follow-up (6–9 months). Associations between cluster membership and follow-up depression severity were examined using linear and logistic regression models adjusted for baseline severity.

Results

The number of clusters varied across folds (3–5). Agreement between solutions was modest (mean ARI = 0.34, range 0.17–0.53; permutation-p = 0.65), indicating limited stability. A consensus clustering approach based on the cross-validation yielded a three-cluster solution (n = 72, 59, 21), with 154/320 patients showing high assignment consistency (≥ 0.75) and classified as stable. Generalizability was low: held-out accuracy averaged 28.1% (17.2–37.5%; permutation-p = 0.39), and assignment confidence was modest (median margin = 0.64, IQR = 0.90; permutation-p = 0.06). No differences were observed between stable clusters in sociodemographic or clinical characteristics, and cluster membership was not associated with treatment outcome at follow-up. In both linear and logistic models, cluster membership was non-significant (p = 0.48; 0.66), whereas baseline severity predicted outcome (p < 0.001; 0.01).

Conclusions

Three stable cognitive clusters were identified, with both cold, hot, and social cognitive domains contributing to cluster profiles. However, the clustering solution showed moderate stability and limited generalizability. Importantly, cluster membership was not associated with treatment outcome at first follow-up, while baseline symptom severity consistently predicted outcome. No differences were observed in sociodemographic or clinical characteristics between clusters. These findings suggest that although cognitive heterogeneity can be detected, the derived subtypes may have limited robustness and clinical utility for predicting treatment response in MDD.

References

1. K. V. Athira, S. Bandopadhyay, P. K. Samudrala, V. G. M. Naidu, M. Lahkar, and S. Chakravarty, “An Overview of the Heterogeneity of Major Depressive Disorder: Current Knowledge and Future Prospective,” Curr. Neuropharmacol., vol. 18, no. 3, pp. 168–187, 2019, doi: 10.2174/1570159x17666191001142934.
2. M. Maj et al., “The clinical characterization of the adult patient with depression aimed at personalization of management,” World Psychiatry, vol. 19, no. 3, pp. 269–293, 2020, doi: 10.1002/wps.20771.
3. R. W. Lam, S. H. Kennedy, R. S. McIntyre, and A. Khullar, “Cognitive Dysfunction in Major Depressive Disorder: Effects on Psychosocial Functioning and Implications for Treatment,” Can. J. Psychiatry, vol. 59, no. 12, pp. 649–654, 2014, doi: 10.1177/070674371405901206.
4. V. H. Dam et al., “Evaluating cognitive disturbances as treatment target and predictor of antidepressant action in major depressive disorder: A NeuroPharm study,” Transl. Psychiatry, vol. 12, no. 1, p. 468, 2022, doi: 10.1038/s41398-022-02240-1.
5. K. H. R. Jensen et al., “Deep phenotyping towards precision psychiatry of first-episode depression — the Brain Drugs-Depression cohort,” Bmc Psychiatry, vol. 23, no. 1, p. 151, 2023, doi: 10.1186/s12888-023-04618-x.