Machine learning for immune biomarkers in severe mental illness: a systematic review

The integration of machine learning (ML) approaches with immune biomarker research may facilitate the identification of candidate markers for achieving personalized medicine approaches in severe mental illnesses (SMI). This systematic review synthesizes the available evidence on ML algorithms applied to immune biomarkers in major depressive (MDD), bipolar (BD) and schizophrenic spectrum disorders (SZ), examining their performance across different clinical uses including diagnostic, prediction, monitoring, prognostic categories, in accordance with the Food and Drug Administration - Biomarker, EndpointS, and other Tools (FDA BEST) framework. We performed a PRISMA-compliant systematic search of PubMed, Web of Science, Scopus and PsycINFO databases until 14 July 2025, including 43 eligible studies with a total sample of 11,556 participants, 8339 with SMI (3228 MDD, 2614 BD and 2497 SZs) and 3217 healthy controls. We systematically described population, ML input data (including blood collection conditions, pre-processing steps, sample type, laboratory assay, missing data, and multimodality), and algorithms (supervised versus unsupervised models, feature selection, validation strategy, outcomes, and performance metrics). Overall, ML models showed moderate to high but heterogeneous performance. Diagnostic applications were the most common (AUC = 0.650-0.990), though predictive, monitoring, and prognostic uses were underrepresented and more variable. Across disorders, pro-inflammatory markers (IL-6, IL-8, TNF-α, IFN-γ, CRP) and IL-10 emerged most consistently, and data-driven approaches suggested shared immune subtypes beyond categorical diagnoses. However, substantial methodological and biological heterogeneity was observed, including inconsistent handling of missing data, limited external validation, and variable feature selection. Immunology-specific sources of variability (such as fasting status, circadian rhythms, and measurement batch effects) were rarely addressed, and the long-term stability of immune-based ML signatures remains largely unexplored. These gaps currently limit clinical translation and underscore the need for standardized protocols and more rigorous ML pipelines.