A Cross Attention Approach to Diagnostic Explainability Using Clinical Practice Guidelines for Depression

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Mental Health Research

Background:

• Explainability is a barrier to AI adoption in analyzing clinical dialogue.
• Online mental health communities offer vast, untapped data for AI analysis.
• Clinician expertise can guide AI in understanding patient interactions.

Purpose of the Study:

• To develop an AI method that leverages clinical knowledge for classifying and explaining depression-related data.
• To enhance transformer models for generating understandable explanations for mental health practitioners (MHPs).
• To improve the trustworthiness and efficiency of AI in mental health screening.

Main Methods:

• Proposed ProcesS knowledgeinfused cross ATtention (PSAT), a domain-general architecture incorporating clinical practice guidelines (CPG).
• Transformed depression-focused CPGs (e.g., PHQ-9) into a machine-readable ontology using SNOMED-CT.
• Enhanced GPT-3.5's ability to generate application-relevant explanations by integrating clinical knowledge.

Main Results:

• PSAT generated application-relevant explanations for depression-related data across four expert-curated datasets.
• PSAT outperformed twelve baseline models in classification and explanation generation.
• PSAT provided explanations where other models failed.

Conclusions:

• PSAT effectively integrates clinical knowledge into AI models for explainable depression analysis.
• The PSAT architecture enhances trust and utility of AI tools for mental health screening.
• This approach has significant potential for both online and offline mental health applications.