Identification of Patients with Acute Lung Injury from Free-Text Chest X-Ray Reports

Identification of complex clinical phenotypes among critically ill patients is a major challenge in clinical research. The overall research goal of our work is to develop automated approaches that accurately identify critical illness phenotypes to prevent the resource intensive manual abstraction approach. In this paper, we describe a text processing method that uses Natural Language Processing (NLP) and supervised text classification methods to identify patients who are positive for Acute Lung Injury (ALI) based on the information available in free-text chest x-ray reports. To increase the classification performance we enhanced the baseline unigram representation with bigram and trigram features, enriched the n-gram features with assertion analysis, and applied statistical feature selection. We used 10-fold cross validation for evaluation and our best performing classifier achieved 81.70% precision (positive predictive value), 75.59% recall (sensitivity), 78.53% f-score, 74.61% negative predictive value, 76.80% specificity in identifying patients with ALI.