DeLLiriuM: A large language model for delirium prediction in the ICU using structured EHR

Area of Science:

• Artificial Intelligence in Medicine
• Critical Care Medicine
• Clinical Informatics

Background:

• Delirium affects up to 31% of intensive care unit (ICU) patients, characterized by fluctuating attention and cognitive impairment.
• Early delirium detection is crucial for timely interventions and improved patient outcomes.
• Existing artificial intelligence (AI) models for ICU delirium prediction often use limited datasets or older AI techniques.

Purpose of the Study:

• To introduce DeLLiriuM, a novel large language model (LLM)-based delirium prediction model for the ICU.
• To utilize structured electronic health record (EHR) data from the first 24 hours of ICU admission for delirium risk prediction.
• To develop and validate a model capable of capturing clinical context from structured EHR data.

Main Methods:

• Developed DeLLiriuM, an LLM-based model, by transforming structured EHR data into an unstructured text format.
• Trained and validated the model on a large, multi-center dataset comprising 104,303 ICU admissions from three databases (eICU, MIMIC-IV, UF-IDR).
• Evaluated model performance using the area under the receiver operating characteristic curve (AUROC) on an external validation set.

Main Results:

• DeLLiriuM demonstrated superior performance compared to baseline models on the external validation set.
• The model achieved an AUROC of 82.5 (95% CI 81.8-83.1) across 77,543 patients from 194 hospitals.
• The LLM approach effectively captured clinical contextual information from transformed EHR data, enhancing predictive accuracy.

Conclusions:

• DeLLiriuM represents the first LLM-based delirium prediction tool for the ICU utilizing structured EHR data.
• The model's ability to process transformed EHR data offers improved delirium risk prediction.
• This approach holds potential for enhancing early detection and management of delirium in critical care settings.