Development and prospective implementation of a large language model based system for early sepsis prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Sepsis, a life-threatening condition, requires early detection for improved patient outcomes.
• Current computational models for sepsis prediction often lack contextual insights from unstructured clinical notes.
• Existing methods struggle to differentiate sepsis from sepsis-mimics, impacting diagnostic accuracy.

Purpose of the Study:

• To introduce COMPOSER-LLM, an open-source large language model (LLM) integrated with the COMPOSER model.
• To enhance early sepsis prediction by leveraging unstructured clinical data.
• To improve the accuracy of sepsis detection by assessing sepsis-mimics using contextual information.

Main Methods:

• Integration of a large language model (LLM) with the existing COMPOSER model.
• Utilizing the LLM to extract contextual information from unstructured clinical notes for high-uncertainty predictions.
• Evaluation of the COMPOSER-LLM model on 2500 patient encounters and prospective validation.

Main Results:

• COMPOSER-LLM demonstrated improved performance over the standalone COMPOSER model.
• Achieved a sensitivity of 72.1%, positive predictive value of 52.9%, and F-1 score of 61.0%.
• Manual chart review indicated potential clinical utility, with 62% of false positives having bacterial infections.

Conclusions:

• Integrating LLMs with traditional models significantly enhances predictive performance for sepsis.
• Leveraging unstructured data through LLMs represents a substantial advancement in healthcare analytics.
• COMPOSER-LLM shows promise for improving early sepsis detection and patient management.