Speaker Role Identification in Clinical Conversations

Area of Science:

• Medical Informatics
• Computational Linguistics
• Artificial Intelligence in Healthcare

Background:

• Patient-clinician communication is vital for understanding healthcare interactions and outcomes.
• Manual analysis of clinical discourse is time-consuming and challenging, particularly for Speaker Role Identification (SRI).
• Existing automatic speech recognition systems with diarization lack the ability to assign specific roles to speakers.

Purpose of the Study:

• To investigate the effectiveness of Large Language Models (LLMs) for Speaker Role Identification (SRI) in clinical settings.
• To evaluate SRI performance using linguistic features alone versus incorporating diarization identifiers.
• To assess the impact of identifier corruption on LLM-based SRI accuracy.

Main Methods:

• Utilized BERT, a Large Language Model, for Speaker Role Identification (SRI) on clinical transcripts.
• Employed veridical turn segmentation and diarization identifiers.
• Fine-tuned the BERT model with varying levels of identifier corruption to test performance robustness.

Main Results:

• BERT achieved 82% accuracy and an 82% F1-score for SRI using only linguistic signals.
• Incorporating accurate diarization identifiers significantly improved performance to 95% accuracy and a 95% F1-score.
• The study demonstrated LLMs' capability in SRI within clinical contexts.

Conclusions:

• Fine-tuned Large Language Models (LLMs) are highly effective for automatic Speaker Role Identification (SRI) in clinical transcripts.
• LLMs offer a robust solution for analyzing patient-clinician communication dynamics.
• Combining LLMs with accurate diarization further enhances the precision of SRI in healthcare settings.