Dissecting first-episode psychosis heterogeneity with clustering analyses: a systematic review.

INTRODUCTION: First episode psychosis (FEP) refers to a variable clinical presentation of the onset of psychotic symptoms and represents a key point in time to ensure prompt treatment of affected patients. Clustering techniques based on unsupervised machine learning (ML) allow the stratification of heterogeneous patient groups to improve diagnostic and prognostic accuracy. This systematic review focuses on the application of unsupervised ML clustering to patients affected by FEP, assessing its potential in clinical decision making and its current limitations. METHODS: A bibliographic search was conducted from inception through December 31, 2025 on PubMed, Embase, Psycinfo and Scopus, following the PRISMA guidelines. RESULTS: 48 studies met the inclusion criteria by showing a partitioning of the sample into clusters according to a cognitive and functional, immune and genetic, clinical, or imaging and neurophysiological dimension and by using one or more unsupervised ML techniques. The immune and genetic and imaging and neurophysiological studies suggest a two-cluster classification, with one subgroup showing a higher inflammatory state or pre-existing brain damage. Overall, no consensus emerged regarding the number or characteristics of clusters identified across cognitive and functional, and clinical domains, due to differences in analysed features and methodological approaches. Nevertheless, a recurring finding is the identification of a subgroup characterized by greater cognitive, functional or clinical impairment already at the onset of psychosis. This suggests that, despite the lack of a consistent classification framework, unsupervised clustering may help identify patients who need a higher therapeutic commitment. CONCLUSIONS: Unsupervised ML clustering for classification of FEP-affected patients has significant potential to improve diagnostic workflow and individualized treatment strategies, but lack of reproducibility remains a major limitation. Future larger studies integrating multi-omics and longitudinal data are warranted.