Identifying Patients With Delirium Based on Unstructured Clinical Notes: Observational Study.

Area of Science:

• Medical Informatics
• Clinical Research
• Artificial Intelligence in Healthcare

Background:

• Delirium is an acute brain dysfunction in hospitalized patients.
• Current diagnostic codes (ICD) for delirium in electronic health records (EHRs) lack accuracy.
• Accurate delirium detection is crucial for research and quality improvement.

Purpose of the Study:

• To develop a more accurate method for detecting delirium episodes using natural language processing (NLP).
• To leverage unstructured clinical notes for improved delirium identification in EHRs.

Main Methods:

• Trained three NLP classifiers (SVM, RNN, Transformer) on 1.5 million clinical notes from over 10,000 patients.
• Utilized expert-labeled sentences for model training and external datasets for testing.
• Evaluated model performance using F1 scores, AUCs, and compared associations with ICD codes, medications, restraints, and mortality.

Main Results:

• The Transformer NLP model achieved high performance (micro F1=0.978, macro F1=0.918).
• NLP detections showed stronger correlations with deliriogenic medications, restraints, mortality, and CAM scores compared to ICD codes.
• NLP demonstrated superior accuracy in identifying delirium indicators.

Conclusions:

• Clinical notes offer a valuable data source for EHR-based delirium studies when analyzed with automated methods.
• The developed NLP model accurately detects delirium, comparable to manual chart review.
• This NLP approach enhances the determination of delirium for large-scale studies, quality improvement initiatives, and clinical trials.