ProKnow: Process knowledge for safety constrained and explainable question generation for mental health diagnostic...

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems
• Natural Language Processing

Background:

• Virtual Mental Health Assistants (VMHAs) currently offer counseling but lack diagnostic capabilities due to safety and clinical process knowledge constraints.
• Existing large-scale language models (LMs) struggle to adhere to specialized clinical workflows required for accurate patient diagnosis.
• Clinical diagnosis necessitates a deep understanding of evidence-based guidelines and expert conceptual knowledge, termed Process Knowledge (ProKnow).

Purpose of the Study:

• To introduce ProKnow-data, a novel dataset of diagnostic conversations incorporating safety constraints and ProKnow.
• To develop ProKnow-algo, a natural language question generation method for interactive diagnostic information collection.
• To evaluate the efficacy of ProKnow-algo in enhancing the safety, knowledge capture, and explainability of diagnostic questions generated by LMs.

Main Methods:

• Developed ProKnow-algo, a method that explicitly models safety, knowledge capture, and explainability for diagnostic question generation.
• Created ProKnow-data, a dataset of clinical diagnostic conversations adhering to safety and ProKnow.
• Involved expert clinicians in designing evaluation metrics focusing on safety, logical coherence, knowledge capture, and explainability.

Main Results:

• LMs utilizing ProKnow-algo generated 89% safer questions in mental health domains.
• ProKnow-algo achieved a 96% reduction in knowledge capture metrics compared to baseline LM generations, ensuring adherence to clinical process knowledge.
• ProKnow-algo demonstrated an average 82% improvement in safety, explainability, and process-guided question generation across various LMs.

Conclusions:

• ProKnow-algo significantly enhances the capability of LMs for diagnostic question generation in healthcare by integrating crucial safety and clinical process knowledge.
• The developed ProKnow-data and ProKnow-algo offer a reproducible framework for advancing AI in clinical diagnostics.
• This work paves the way for more reliable and safer AI-driven diagnostic support in mental healthcare settings.