Machine Learning-Assisted Identification of Biomarkers Associated with Schizotypal and Cyclothymic Traits

Introduction: Schizotypal traits (unusual perceptual experiences and cognitive-perceptual differences) and cyclothymic traits (mood instability and affective fluctuations) represent subclinical personality dimensions observed in non-clinical populations. Allostatic load (AL), reflecting cumulative multisystem physiological stress, may provide insight into biological processes associated with these traits. This study examined whether individuals with elevated schizotypy or cyclothymic traits show distinct AL-related biomarker patterns. Methods: Three groups were defined using validated psychometric measures: 30 individuals with schizotypy traits (PSF), 25 with cyclothymic traits (CF), and 30 control participants (CGs). Machine learning (ML) analyses included a Random Forest model with leave-one-subject-out cross-validation to examine differences in biomarker profiles between groups, and Shapley Additive exPlanations (SHAP) analysis was applied to determine the relative importance of biomarkers. The highest-ranking biomarkers were subsequently used to construct a reduced AL index. Results: Significant differences among groups were observed in oxidative stress system markers (p = 0.014). SHAP analysis identified creatinine, diastolic blood pressure, uric acid, and low-density lipids as key features for the CF group, and heart rate, TSH, uric acid, and glucose for the PSF group, with uric acid emerging as the strongest shared biomarker. A reduced biomarker panel selected using SHAP achieved subject-level classification accuracy of 63.6% (95% CI 50.9–76.4) and AUC of 0.625 (95% CI 0.469–0.768) and showed comparable performance to the full model. However, label-permutation testing yielded a borderline result (p = 0.055), indicating that these findings should be considered exploratory. The AL index derived from the top-ranked biomarkers was significantly lower in the CF than in the CG group (p = 0.007), whereas differences between the PSF and CG groups did not remain significant after correction. Findings are exploratory and require validation in an independent cohort. Conclusions: Distinct biomarker patterns associated with schizotypal and cyclothymic traits were identified in a non-clinical population, highlighting differences in stress-related physiological processes. ML-based feature-importance analysis identified a reduced set of biomarkers associated with trait-related biological profiles; however, these findings remain exploratory and require external validation.