Large language models in respiratory care: safety and clinical integration.

BACKGROUND

Large language models (LLMs) are rapidly entering respiratory medicine workflows. Their clinical role remains unclear. A central concern is whether they function as autonomous decision-makers or as tools that depend on clinician input.

METHODS

We conducted a structured narrative review across three domains aligned with respiratory practice, including bedside clinical decision support and triage, pulmonary function test (PFT)/spirometry interpretation, and chest radiograph (CXR) reporting. We included quantitative studies comparing LLMs or specialised artificial intelligence (AI) systems with clinician performance and extracted accuracy, agreement and safety-critical failure modes.

RESULTS

At bedside, LLMs approached expert diagnosis in straightforward cases, but performance fell in highly complex or multi-step management tasks. The dominant error was omission of required actions, unsafe under-triage or incomplete plans. In PFTs, guideline-anchored and spirogram-trained LLMs achieved high agreement and reduced inter-observer variability on structured inputs. In contrast, generalist LLMs showed only moderate reliability and inconsistent handling of borderline or mixed patterns. In CXR reporting, chest-specific multimodal systems outperformed generalist LLMs, with fewer hallucinations and better clinical acceptability; yet clinically relevant discrepancies persisted without radiologist review.

CONCLUSIONS

Across domains, AI does not replace respiratory expertise. When inputs are complete, structured and guideline-concordant, outputs stabilise and add value. When inputs are ambiguous or incomplete, errors expand, often through dangerous omissions. Safe deployment therefore requires specialised training, strict input structure and continuous clinician oversight.