Artificial intelligence in emergency mechanical ventilation: a prospective observational comparison with emergency...

Artificial intelligence (AI) has the potential to support clinicians in high-risk and complex decision-making processes, such as mechanical ventilation. This prospective observational study aimed to compare mechanical ventilator settings determined by emergency physician (EP) with recommendations generated by three large language models (ChatGPT-5, Gemini, and Copilot) in the emergency department (ED).

This prospective, analytical, single-center study included 30 intubated patients managed in an ED over a three-month period. Clinical data, including diagnoses, vital signs, and initial arterial blood gas parameters, were presented to ChatGPT-5, Gemini, and Copilot. The AI models' recommendations for ventilation mode, tidal volume, respiratory rate, PEEP, and FiO₂ were compared with the initial settings adjusted by EP. Agreement for ventilator mode selection was assessed using Cohen's kappa statistics, while agreement for continuous ventilator parameters was evaluated using Bland-Altman analysis.

A total of 30 patients were included in the study. The median age was 73 years (IQR: 60-84), and 66.7% were male. When the ventilator setting preferences of the EP were analyzed, the most commonly used modes were VCV (46.7%) and SIMV (40.0%). Among the AI models, ChatGPT-5 primarily recommended VCV (76.7%) and, to a lesser extent, CPAP (10.0%); Gemini most frequently preferred VCV (56.7%) and PCV (43.3%); and Copilot predominantly recommended PCV (70.0%). Data on the compatibility of mechanical ventilator mode selection revealed that AI models showed 'poor' agreement with expert opinion (EP) based on diagnosis. ChatGPT showed 50.0% agreement (Cohen's kappa: 0.199; 95% Confidence Interval (CI): -0.087 to 0.486), Google Gemini 43.3% agreement (Cohen's kappa: 0.164; 95% CI: -0.098 to 0.426), and Microsoft Copilot 20.0% agreement (Cohen's kappa: -0.043; 95% CI: -0.230 to 0.143).

Agreement between AI-generated ventilator settings and the EP was limited. Current AI models may offer supportive input; however, these findings should be interpreted as preliminary and exploratory, and further large-scale, multicenter studies are needed to validate these results.

Not applicable.