Symptom-based phenotypes in recent-onset psychosis: derivation and neuroimaging validation through multimodal machine learning

Introduction: Schizophrenia and related psychoses show substantial clinical heterogeneity, with 30% to 50% of patients responding inadequately to treatment. This variability reflects complex, multiscale interactions of molecular, structural and functional alterations with symptom expression, which current diagnostic and therapeutic approaches fail to accommodate [1].

A critical step towards precision care is the identification of clinically and biologically meaningful subgroups within psychosis. Current clinical scales capturing broader symptom dimensions may overlook a finer-grained structure of psychopathology[2]. Data-driven approaches are needed for more nuanced symptom dimensions and their relevance to neurobiological features may advance targeted interventions.

Objective: To derive robust symptom-based phenotypes in recent-onset psychosis (ROP) and evaluate their neurobiological underpinnings through multimodal neuroimaging-based classification.

Methods: We included 454 ROP patients from the MUNICH (n = 139) and PRONIA (n = 315, across 10 sites) cohorts (mean age = 27.1 ± 7.5 years; 286 males). Symptom dimensions were derived using Orthogonal Projective Non-Negative Matrix Factorization (OPNMF) applied to the Positive and Negative Syndrome Scale [3] and the Scale for the Assessment of Negative Symptoms [4] ratings. We evaluated 3-, 4-, and 5-factor solutions. The resulting component matrix was projected onto the PRONIA sample, with patients assigned to the factor with their highest loading.

We classified PRONIA patients according to their assigned factors against matched healthy controls (HC) (n=522, mean age = 25.6 ± 6.0 years; 306 females). Neuroimaging features included gray matter volume (GMV) derived from T1-weighted images and fractional amplitude of low-frequency fluctuations (fALFF) derived from resting-state fMRI in three frequency sub-bands (slow-3 to slow-5). Images were corrected for age, sex and site using dynamic standardization for GMV and offset correction for fALLF features. Classification was performed using a linear support vector machine within a nested cross validation framework (inner/outer cross validation: 5 permutations x 10 folds). Features underwent ranking, thresholding, minimum redundancy maximum relevance-based selection, dimensionality reduction, and standardization before training.

Results: The 4-factor solution outperformed others, showing the highest explained variance in both the combined (89.3%) and PRONIA (88.0%) sample, as well as the strongest neuroimaging-based predictive accuracy. This solution yielded four clinically interpretable phenotypes: Avolition-Asociality, Expressive Deficits, Cognitive Disorganization, and Positive Symptoms, with a balanced distribution of patients across subgroups.

Classification of patients assigned to factors from matched HC using GMV and fALFF features showed highest performance with the 4-factor model and slow-3 fALFF data (F1: sensitivity (sens) = 65.22%, specificity (spec) = 64.49% , balanced accuracy (BAC)= 64.85%, area under the curve (AUC) = 0.65 (0.57-0.74), F2: sens = 63.04%, spec= 68.18%, BAC= 65.61%, AUC= 0.72 (0.62-0.81), F3: sens= 83.67%, spec=63.86%, BAC=73.76%, AUC=0.8 (0.72-0.89), F4: sens=55.17%, spec=73.97%, BAC=64.57%, AUC=0.7 (0.63-0.77).

Conclusion: We identified four interpretable symptom dimensions in ROP patients, extended beyond traditional clinical subscales. These phenotypes were differentiable from HC using multimodal neuroimaging features. Our findings support the idea that richer symptom dimensions can reveal biologically grounded subgroups in psychosis, offering a foundation for more personalized clinical approaches.