Automated speech-fluency explanations for schizophrenia diagnosis

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Clinical Psychology

Background:

• Schizophrenia diagnosis relies on time-intensive clinician assessments.
• Existing automated diagnostic tools lack crucial explainability for clinical use.
• Speech analysis offers a potential avenue for objective schizophrenia assessment.

Purpose of the Study:

• To develop a fully automated and explainable pipeline for schizophrenia detection from audio recordings.
• To leverage automatic speech recognition (ASR) and large language models (LLMs) for enhanced diagnostic capabilities.
• To provide transparent AI-driven insights for clinical decision-making in schizophrenia diagnosis.

Main Methods:

• Evaluated three ASR models (Truebar, Whisper, Soniox) for transcription accuracy.
• Utilized LLMs to semantically enrich speech transcriptions and extract verbal/non-verbal features.
• Employed a Bayesian framework for feature relevance assessment and trained machine learning models for classification.

Main Results:

• The best model, an Explainable Boosting Machine, achieved 0.82 classification accuracy and 0.90 AUC.
• Generated visual explanations for model predictions, enhancing interpretability.
• Established the first automated and explainable schizophrenia detection framework for the Slovene language.

Conclusions:

• The developed pipeline offers a novel, explainable approach to schizophrenia detection using speech data.
• The system provides performance comparable to existing automated methods while prioritizing transparency.
• This framework supports informed clinical decision-making through understandable AI outputs.