Evaluating Report Text Variation and Informativeness: Natural Language Processing of CT Chest Imaging for Pulmonary Embolism

Objective: The aim of this study was to quantify the variability of language in free text reports of pulmonary embolus (PE) studies and to gauge the informativeness of free text to predict PE diagnosis using machine learning as proxy for human understanding. Materials and Methods: All 1,133 consecutive chest CTs with contrast studies performed under a PE protocol and ordered in the emergency department in 2016 were selected from our departmental electronic workflow system. We used commercial text-mining and predictive analytics software to parse and describe all report text and to generate a suite of machine learning rules that sought to predict the "gold standard" radiological diagnosis of PE. Results: There was extensive variation in the length of Findings section and Impression section texts across the reports, only marginally associated with a positive PE diagnosis. A marked concentration of terms was found: for example, 20 words were used in the Findings section of 93% of the reports, and 896 of 2,296 distinct words were each used in only one report's Impression section. In the validation set, machine learning rules had perfect sensitivity but imperfect specificity, a low positive predictive value of 73%, and a misclassification rate of 3%. Conclusion: Use of free text reporting was associated with extensive variability in report length and report terms used. Interpretation of the free text was a difficult machine learning task and suggests potential difficulty for human recipients in fully understanding such reports. These results support the prospective assessment of the impact of a fully structured report template with at least some mandatory discrete fields on ease of use of reports and their understanding.