# From the p-Factor to Cognitive Content: Detection and Discrimination of Psychopathologies Based on Explainable Artificial Intelligence

## TL;DR

This study uses AI to detect and differentiate mental health conditions like depression and anxiety based on language patterns, offering new insights into shared and distinct features of psychopathologies.

## Contribution

The study introduces a novel approach combining NLP and ML with cognitive content analysis to detect and differentiate psychopathologies using linguistic markers.

## Key findings

- Anxiety was marked by positive content and hope, while depression involved negative and hopeless language.
- Depressive-anxiety combined anxiety features with a negative outlook, suggesting a transitional phase.
- Models showed high accuracy in distinguishing self-reported pathology diagnoses from subclinical samples.

## Abstract

Background and Aims: Differentiating psychopathologies is challenging due to shared underlying mechanisms, such as the p-factor. Nevertheless, recent methodological advances suggest that distinct linguistic markers can help detect and differentiate these conditions. This study aimed to use cognitive content analysis with advanced natural language processing (NLP) and machine learning (ML) to (Study 1) distinguish among control, depression, anxiety, and depressive-anxiety groups and (Study 2) detect general psychopathology.

Methods: Data from 1901 participants (retained from 2551 respondents aged 18–43 years who completed the Beier sentence completion test [BSCT]) were analyzed. For Study 1, groups were formed using the Depression, Anxiety, and Stress Scale (DASS-21); negative mood was assessed via the Positive and Negative Affect Schedule (PANAS). For Study 2, the Brief Symptom Inventory (BSI) categorized general psychopathology and self-reported diagnostic status served as external validation. Two analytic approaches were employed: (1) textual analysis with a bidirectional encoder representations from transformers (BERT) model and (2) subscale-score analysis using a support vector machine (SVM). SHapley Additive exPlanations (SHAP) interpreted the ML models.

Results: In Study 1, the models distinguished control, depression, anxiety, and depressive-anxiety groups. Anxiety was marked by positive content, hope, and I-Talk, whereas depression involved negative, hopeless content. Depressive-anxiety combined features of anxiety with a pronounced negative outlook, suggesting a transitional phase where diminishing hope may bridge anxiety to depression. In Study 2, the models performed high in distinguishing the self-reported pathology diagnosis group (area under the curve [AUC]: 0.81 [BERT], 0.85 [SVM]) from subclinical samples but failed to differentiate the self-reported past diagnosis (AUC: 0.53 [BERT], 0.57 [SVM]) group from controls. This implies that cognitive changes in psychopathology may share a consistent underlying structure like p-factor.

Conclusion: These pioneer findings demonstrate that integrating advanced computational techniques can identify key linguistic markers and guide the development of language-based diagnostic tools, potentially transforming mental health diagnostics.