Phenomenological psychopathology meets machine learning: A multicentric retrospective study (Mu.St.A.R.D.) targeting the role of Aberrant Salience assessment in psychosis detection.

BACKGROUND

The Aberrant Salience (AS) model conceptualizes psychosis onset as the altered attribution of salience to neutral stimuli. The Aberrant Salience Inventory (ASI), a psychometric tool, measures this phenomenon. This study utilized a multi-center, multi-country retrospective dataset to refine the ASI's screening capabilities using decision tree (DT) and multilayer perceptron (MLP) models. By integrating machine learning approaches with practical screening utility and phenomenological insights, this work advances understanding of psychosis development and its mechanisms.

METHODS

Data from 2981 individuals (537 people with a psychosis diagnosis, 2444 controls) were analyzed, including ASI item responses, age, and sex. Classification models - DTs and MLPs - were applied with hyperparameter tuning and cost adjustments for imbalanced data. Least Absolute Shrinkage and Selection Operator (LASSO) feature selection identified key ASI items for DT construction, and separate DTs were computed for demographic subgroups.

RESULTS

The psychosis group exhibited higher mean ASI scores and older average age than controls. The seven-item DT model achieved a balanced accuracy of 67.4 % and AUC of 0.72, with enhanced performance in older individuals and males. The MLP model outperformed the DT (balanced accuracy: 73.28 %, AUC: 0.74).

CONCLUSIONS

AS emerges as a critical paradigm in psychosis psychopathology, complementing other psychosis staging models. The seven-item DT demonstrated strong clinical utility, reflecting key psychopathological constructs like delusional mood and apophany. Future studies should validate findings in diverse populations and explore integration with neurobiological markers for refined psychosis prediction.