RespBERT: A Multi-Site Validation of a Natural Language Processing Algorithm, of Radiology Notes to Identify Acute...

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Radiology

Background:

• Acute Respiratory Distress Syndrome (ARDS) is a severe condition in critically ill patients with high mortality.
• Current ARDS diagnosis relies on the Berlin criteria, often requiring manual review of chest radiographs.
• Electronic Medical Records (EMRs) frequently lack explicit radiological findings needed for ARDS diagnosis.

Purpose of the Study:

• To develop and evaluate an automated Natural Language Processing (NLP) pipeline for identifying potential ARDS cases.
• To analyze radiology reports for radiological evidence of ARDS using machine learning.
• To facilitate large-scale studies of ARDS by automating the identification of cases from EMRs.

Main Methods:

• A Natural Language Processing (NLP) pipeline was created to analyze radiology notes from 362 patients meeting sepsis-3 criteria.
• A BERT-based classification model was employed to identify features indicative of ARDS.
• The model was trained and validated on two distinct datasets (Emory and Grady).

Main Results:

• The NLP pipeline demonstrated effectiveness in classifying patients for possible ARDS diagnosis.
• Classification models achieved F1-scores of 74.5% for the Emory dataset and 64.22% for the Grady dataset.
• Automated analysis of radiological reports shows promise for ARDS identification.

Conclusions:

• Automated analysis of radiology reports using NLP can aid in the identification of ARDS.
• This approach can overcome limitations of manual review and EMR data availability.
• Further development of NLP tools can enhance ARDS research and clinical management.