Data-driven symptom phenotypes in psychosis: derivation and neuroimaging validation through machine learning

Introduction: Psychotic disorders are characterized by substantial clinical heterogeneity, posing a significant challenge in the development of effective, personalized treatments. The Virtual Brain Twin project seeks to address this challenge by developing personalized, mechanistic brain models that integrate clinical, imaging, and molecular data to guide treatment decisions in psychosis. Symptom-based phenotypes can provide an “initial coordinate system” to reduce this heterogeneity into more homogeneous disease dimensions and for linking distinct symptom profiles to underlying pathophysiological mechanisms.

Aim: The objective of this study is to derive robust symptom-based phenotypes in psychosis and evaluate their neurobiological underpinnings through multimodal neuroimaging-based classification.

Methods: The study pooled clinical and neuroimaging data from three cohorts with patients with recent onset psychosis and established phases of schizophrenia: PRONIA (N=359, mean age 25.7±5.8), MUC (N=139, mean age 30.3±9.4), and MIMICSS (N=70, mean age 35.0±12.0), MUC (N=139, mean age 30.3±9.4). Symptom dimensions were derived using Orthogonal Projective Non-Negative Matrix Factorization (OPNMF) applied to PANSS and SANS items. Solution stability was evaluated across 1,000 resampling iterations (bootstrap, half-split, and cross-validation) using four complementary metrics: Jaccard Index, Adjusted Rand Index, normalized Variation of Information and Concordance Index. Generalizability was assessed via transferred reconstruction error. Factor-specific patient subgroups were classified against matched healthy controls using a Linear Support Vector Machine trained on gray matter volume (GMV) within a nested cross validation framework. To prevent information leakage between the label-generator and the prediction model, factorization was embedded within an external layer of cross-validation (CV3, 5 folds). Performance was assessed both in the CV3 training and test samples. Neuroimaging features were characterized using Cross-Validation Ratio (CVR) maps masked by Sign-Based Consistency-derived significance maps (FDR-corrected, p < 0.05). The masked CVR maps were averaged over the CV3 folds and thresholded at |CVR|>2 to reveal factor-specific significant features.

Results Stability analysis identified a three-factor solution as optimal across all stability metrics, with generalizability showing a linear improvement with increasing number of factors. The three symptom dimensions were: F1 – Expressive Deficits, F2 – Positive Symptoms, and F3 – Avolition-Asociality. Patients were successfully classified against healthy controls using GMV features, with training CV3 fold BAC exceeding test fold performance, reflecting realistic generalization to unseen data. Stable neuroimaging patterns (surviving the full CVR and frequency thresholding pipeline) were identified for F2 and F3 with no stable voxels surviving for F1. For F2 gray matter reductions (negative CVR) included auditory and peri-Sylvian cortices, multiple thalamic nuclei, and limbic regions (hippocampus, amygdala, cingulate), with a pattern of volume increases (positive CVR) in cerebellar, vermal, and striatal regions. In F3 gray matter reductions were found in a bilateral peri-Sylvian and temporo-parietal network (particularly auditory, sensorimotor, and association cortices) with additional involvement of occipito-temporal and medial frontal regions.

Conclusions This study identified three generalizable symptom-based phenotypes in recent-onset psychosis. Two of the three dimensions showed reproducible neurobiological correlates in gray matter volume, supporting their pathophysiological validity. These phenotypes provide a structured dimensional coordinate system for parsing clinical heterogeneity in psychosis, and serve as a foundation for mechanistic personalized brain modeling within the Virtual Brain Twin project.

References

No references