Data-driven biological subtypes in schizophrenia spectrum disorders and bipolar disorder: a systematic review

Schizophrenia spectrum disorders (SSD) exhibit heterogeneity across clinical presentation, illness trajectories, and neurobiology, alongside clinical and biological overlap with bipolar disorder (BD). The field of biological psychiatry has attempted to disentangle this heterogeneity by applying data-driven clustering to identify subtypes with potentially distinct etiology, prognosis, or treatment needs, a key step toward precision psychiatry. However, the field lacks a comprehensive synthesis of existing evidence.

We conducted a systematic review summarising data-driven clustering studies deriving biological subtypes across SSD and BD. Extracted items included sample demographics, clustering methodology, outcomes, and downstream comparisons (between-cluster and vs. healthy controls). 78 studies representing 82 samples and 92 clustering analyses met inclusion criteria (n = 62 SSD, n = 13 SSD-BD, n = 7 BD samples). Analyses covered structural (n = 27) and functional neuroimaging (n = 8), molecular (n = 11) and peripheral biomarkers (n = 10), electrophysiology (n = 13), multimodal studies (n = 19), and other modalities (n = 4).

Two- to three-subtype solutions predominated, though limited SSD-BD and BD analyses preclude firm conclusions. Evidence was inconsistent for symptomatic or cognitive differences between biological subtypes. Findings were mixed on whether SSD and BD form biologically distinct clusters, with many cross-diagnostic studies not testing diagnostic separation. For inflammatory subtypes in SSD, most studies identified two subtypes, yet a simple high‑ versus low‑inflammation dichotomy was not always supported. Several studies reported subtype differences in treatment response, underscoring the potential of biological subtyping for treatment stratification. A relatively robust finding was the separation of schizophrenia into two neuroanatomical subtypes (one with subcortical volume reductions and another with basal ganglia increases)  showing differential treatment response to antipsychotic medication, potentially linked to dopaminergic mechanisms.

Clustering analyses displayed substantial methodological heterogeneity. The most common clustering algorithms were center-based (e.g., k-means) and hierarchical. 88% of analyses applied dimensionality reduction,  79% addressed cluster number selection, 65% conducted robustness analyses,  while only 33%  tested the null hypothesis of no cluster structure and 37% replicated or validated their clustering solution.

Across 80 univariate tests investigating associations between study outcomes and sample or methodological characteristics, one survived correction: symptomatic differences and sMRI atlas. Other nominally significant associations included cognitive differences and illness duration, number of subtypes and age and fMRI/EEG paradigm , basal ganglia subtype and sMRI submodality, symptomatic differences and study modality , sMRI software, and clustering algorithm.

A co-occurrence matrix was constructed to identify phenotypic features that consistently co-occur within biological subtypes identified across independent studies. This highlighted the co-occurrence of elevated total, positive, negative, and general symptoms and cognitive impairment; inflammatory markers; and neuroanatomical alterations.

By summarizing existing literature, this review provides a structured overview of the field, highlighting recurrent findings, dominant analytical approaches, and gaps in methodological rigor.  Future studies should prioritize rigorous and transparent study designs that explicitly test cluster structure, assess robustness, and incorporate validation or replication, extending to underrepresented modalities, and including sensitivity analyses to quantify the impact of sample and analytic choices.

References

No references