Large language models in respiratory care: safety and clinical integration

Area of Science:

• Artificial intelligence in medicine
• Respiratory medicine applications
• Clinical decision support systems

Background:

• Large language models (LLMs) are increasingly integrated into respiratory medicine workflows.
• The precise clinical role of LLMs remains undefined, with concerns about their autonomy versus tool-based function.

Purpose of the Study:

• To review the performance of LLMs and specialized AI in respiratory medicine.
• To compare AI systems with clinician performance in key respiratory domains.
• To identify accuracy, agreement, and safety-critical failure modes.

Main Methods:

• Structured narrative review across bedside decision support, PFT interpretation, and CXR reporting.
• Inclusion of quantitative studies comparing LLMs/AI with clinician performance.
• Extraction of accuracy, agreement, and safety-critical failure data.

Main Results:

• LLMs approached expert diagnosis in simple bedside cases but struggled with complex tasks, often omitting actions or under-triaging.
• Guideline-anchored LLMs showed high agreement in PFT interpretation, reducing variability.
• Chest-specific AI outperformed general LLMs in CXR reporting, but discrepancies remained without radiologist review.

Conclusions:

• AI does not substitute for respiratory expertise; structured, complete inputs enhance value.
• Ambiguous or incomplete inputs increase AI error rates, particularly dangerous omissions.
• Safe LLM deployment necessitates specialized training, structured inputs, and continuous clinician oversight.