Identifying encephalopathy in patients admitted to an intensive care unit: Going beyond structured information using...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Neurology

Background:

• Encephalopathy is a severe neurological condition in critically ill patients, often under-recorded in electronic health records.
• Underdiagnosis limits large-scale retrospective studies on encephalopathy and its impact.

Purpose of the Study:

• To develop and validate a machine learning model for identifying encephalopathy patients using clinical notes.
• To uncover patients with potential encephalopathy missed in structured electronic health record data.

Main Methods:

• Utilized the MIMIC-III dataset, defining cohorts using ICD-9 codes and clinical concepts.
• Applied Natural Language Processing (NLP) with MedCAT for note annotation and vectorized features.
• Trained Support Vector Machine (SVM) and Random Forest models to classify patients based on clinical notes.

Main Results:

• The best SVM model achieved an 85% F1 score, identifying 31% of patients with possible encephalopathy (clinical notes only) as having the condition with >90% probability.
• Patients identified with possible encephalopathy showed higher length of stay, mortality, and comorbidity rates compared to controls.
• Model validation demonstrated high precision (92%-98%) in identifying new cases.

Conclusions:

• NLP techniques effectively extract crucial clinical information for identifying under-recognized conditions like encephalopathy.
• This machine learning approach successfully identifies numerous patients lacking formal diagnoses in structured EHR data.
• The findings highlight the potential for NLP to improve the diagnosis and study of encephalopathy.