A Computable Phenotype for Acute Respiratory Distress Syndrome Using Natural Language Processing and Machine Learning

Area of Science:

• Medical Informatics
• Computational Linguistics
• Radiology

Background:

• Acute Respiratory Distress Syndrome (ARDS) is a critical form of respiratory failure.
• Identifying ARDS from unstructured radiology reports presents a significant challenge.

Purpose of the Study:

• To compare the performance of Natural Language Processing (NLP) with machine learning against traditional keyword models for ARDS identification.
• To develop and evaluate a computable phenotype for ARDS detection.

Main Methods:

• Linguistic pre-processing of 9,255 radiology reports from 533 patients.
• Machine learning classifiers trained with text features using 10-fold cross-validation.
• Comparison of NLP model performance against a traditional keyword-based model.

Main Results:

• The traditional keyword model achieved 67.3% accuracy and 41.7% positive predictive value (PPV).
• The best NLP model demonstrated significantly higher performance with 83.0% accuracy and 71.4% PPV.
• NLP models showed superior capability in identifying ARDS cases.

Conclusions:

• NLP combined with machine learning offers a more effective approach for ARDS identification in clinical text.
• This advanced method has the potential to increase the detection rate of ARDS cases.
• A computable phenotype for ARDS using NLP can enhance diagnostic capabilities.