DSM Disorders Disappear in Statistical Clustering of Psychiatric Symptoms (2024) (www.psychiatrymargins.com)

A large, data-driven reanalysis of DSM-5 symptoms by Forbes et al. used a demographically diverse online sample (N=14.8K) and 680 independently randomized, first-person symptom items to map the quantitative structure of psychopathology. Responses were clustered with two complementary methods (iclust and Ward’s hierarchical agglomerative clustering), validated across primary and hold-out samples, and organized with hierarchical PCA. The analysis yielded 139 multi-item “syndromes,” 81 solo symptoms, 27 subfactors and eight higher-order spectra (Externalizing; Harmful Substance Use; Mania/Low Detachment; Thought Disorder; Somatoform; Eating Pathology; Internalizing; Neurodevelopmental/Cognitive Difficulties), plus a broad general factor (“Big Everything”). Crucially, classic DSM categories such as Major Depressive Disorder, GAD and PTSD did not emerge as single, statistically homogeneous clusters; instead their diagnostic symptom sets split into smaller, more coherent elements (e.g., depressed mood/anhedonia, self-derogation, suicidality, dysregulated sleep) that cross-cut spectra. For AI/ML and psychiatry researchers this underscores that categorical labels in DSM may mask underlying symptom heterogeneity and comorbidity patterns—bearing on supervised labels, phenotyping for biomarker discovery, and model generalization. The study supports dimensional frameworks like HiTOP and argues for symptom-level, hierarchical representations in predictive models and clinical decision tools. Limitations include reliance on self-report, decontextualized 12-month items and absence of clinician-observed signs; replication with multi-method, longitudinal and more diverse samples will be essential before retooling diagnostic algorithms or clinical practice.