Leveraging large language models for patient-ventilator asynchrony detection.

Area of Science:

• Medical Artificial Intelligence
• Clinical Decision Support Systems
• Respiratory Mechanics

Background:

• Patient-ventilator asynchrony, specifically flow starvation (FS), is a critical issue in mechanical ventilation.
• Accurate detection of FS is essential for optimizing patient outcomes and ventilator management.
• Current methods for FS detection often rely on complex, expert-developed models.

Purpose of the Study:

• To evaluate the performance of large language models (LLMs) in detecting flow starvation (FS) asynchronies.
• To compare LLM performance against expert-developed deep neural networks.
• To assess LLMs' ability to generate code for deep learning models used in FS detection.

Main Methods:

• Four popular LLMs (GPT-4, Claude-3.5, Gemini-1.5, DeepSeek-R1) were tested on 6500 airway pressure cycles.
• LLMs performed direct classification of breaths into three FS categories.
• LLMs generated executable code for one-dimensional convolutional neural network (CNN-1D) and Long Short-Term Memory (LSTM) models.

Main Results:

• LLMs demonstrated poor performance in direct FS classification, with accuracies ranging from 0.497 to 0.627.
• Claude-3.5-generated CNN-1D code achieved a high accuracy of 0.902, surpassing expert-developed models.
• LLMs showed significant potential in automated model development for clinical applications.

Conclusions:

• LLMs have limited direct utility for FS classification but are highly effective in generating deep learning models.
• The ability of LLMs to accelerate the development of clinical tools, like those for detecting patient-ventilator asynchronies, is promising.
• Further validation and ethical considerations are necessary before clinical implementation of LLM-generated models.