Conversational speech for respiratory triage in primary care: a pilot study

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Speech Processing

Background:

• Respiratory complaints are a major reason for adult ambulatory visits.
• Accurate triage is crucial for antibiotic stewardship and targeted therapy.
• Previous voice analysis research focused on single conditions using scripted speech, not real-world primary care conversations.

Purpose of the Study:

• To evaluate the effectiveness of using acoustic features from conversational primary care audio for respiratory illness triage.
• To develop and test voice-based classifiers for differentiating various respiratory conditions.
• To establish a baseline for voice-based clinical AI in differential diagnosis.

Main Methods:

• Utilized a dataset of 514,377 ambient-recorded primary care visits.
• Extracted spectral, voice quality, and prosodic features from doctor-patient conversations.
• Trained 11 binary acoustic classifiers for respiratory triage tasks and compared them against demographic/geographic baselines.

Main Results:

• Acoustic models achieved moderate performance (AUC 0.602-0.745) in differentiating respiratory conditions.
• Six of eleven classifiers outperformed non-acoustic baselines.
• Voice analysis demonstrated discrimination beyond demographic or geographic factors in several tasks.

Conclusions:

• Conversational primary care audio contains acoustic signals that can differentiate respiratory conditions.
• This pilot study demonstrates the feasibility of voice-based differential diagnosis in a primary care setting.
• The findings advance voice-based clinical AI from simple detection to complex diagnostic tasks.