Towards improved specificity in mental health syndromes: projection-based clustering of depressive phenotypes

Background: Quantitative attempts to improve syndrome specificity typically produce large heterogenous sub-groupings, impacting the validity of treatment and research targets. Assessing barriers to the valid application of existing methods and examining improvements from an interpretable projection-based clustering alternative may improve the precision and reproducibility of our research targets and classification systems. Methods: This exploratory, cross-sectional, study recruited 2820 participants aged 12-to-25 years, from primary-healthcare services in Australia, between November 2018 and July 2023. 1843 participants completed relevant self-reported measures of depression, anxiety and mania-like experiences, and were included for analysis. Principal Component Analysis (PCA) was used to examine the distribution of within-syndrome variance. Projection-based subtypes were compared to traditional quantitative phenotyping approaches: clustering paradigms (Model-Based, Centre-Based Partition, Hierarchical), LCA, and Exploratory Factor Analysis (FA). Results: Interpretable projection-based clustering improved homogeneity and qualitative distinctions between clusters were compared to all other methods. This identified 14 clusters organisable into six novel symptom profiles: sleep (n = 117; 11 %), mania (n = 125; 12 %), anxiety (n = 119; 11 %), weight/appetite gain (n = 138; 13 %), weight/appetite loss (n = 242; 23 %), and an undifferentiated type (n = 310; 29 %). The PCA identified a skewed power-law distribution underlying symptom variance, affecting standard LCA/clustering procedures. This interacted with optimisation algorithms, producing heterogenous subtypes. Within FA, it produced the nested hierarchical structure identified in HiTOP studies. Conclusions: A skewed variance distribution underlying the depressive syndrome adversely impacts standard Clustering/LCA methods, and may contribute to past difficulties in identifying well-specified data-driven phenotypes. Consequently, future studies should consider the distribution's impact to their optimisation algorithms or use the better-specified projection-based clustering. Identified profiles reflect major trends in depressive symptom expression, potentially representing improved research targets.