Classification of CT pulmonary angiography reports by presence, chronicity, and location of pulmonary embolism with natural language processing

Abstract

In this paper we describe an efficient tool based on natural language processing for classifying the detail state of pulmonary embolism (PE) recorded in CT pulmonary angiography reports. The classification tasks include: PE present vs. absent, acute PE vs. others, central PE vs. others, and sub-segmental PE vs. others. Statistical learning algorithms were trained with features extracted using the NLP tool and gold standard labels obtained via chart review from two radiologists. The areas under the receiver operating characteristic curves (AUG) for the four tasks were 0.998, 0.945, 0.987, and 0.986, respectively. We compared our classifiers with bag-of-words Naive Bayes classifiers, a standard text mining technology, which gave AUG 0.942, 0.765, 0.766, and 0.712, respectively.

Natural Language Processing Technologies in Radiology Research and Clinical Applications

“…Instead, the major shortcomings are ambiguity, wherein an expression can be interpreted in two or more distinct ways depending on context ( Fig 6); incorrect grammar usage in a fast-paced environment; and misspellings. Despite these challenges, linguistic NLP systems theoretically offer more complete information than pattern matching-based systems and have thus been preferred whenever complex (eg, temporal, anatomic) relationships are of interest (15).…”

“…Although n-grams are powerful features used in many practical systems such as speech recognition, concepts identified by NLP can be more predictive as features, since NLP reduces synonymous findings to standardized names (15). However, this is not always the case because words that are relevant for some characterizations may not have concepts, such as the words "if" and "further," which have been found to be highly predictive for identifying reports with follow-up recommendations (24), thus missing important features.…”

“…This simple system had high sensitivity (86%-98%) and specificity (89%-93%) for each task except chronicity (60% and 99%, respectively) (38). Yu et al (15) used an NLP system called Narrative Information Linear Extraction (NILE) that combines linguistic and machine learning approaches to improve identification of pulmonary embolism location. Although many full-linguistic NLP packages such as MedLEE and cTAKES can identify the anatomy mentioned with a finding, pulmonary embolism presents a particular challenge because of the relevance of the inherently nested anatomic structure of the vasculature (Fig 9).…”

“…a machine learning classifier, the system was able to achieve a receiver operating characteristic area under the curve of 0.998 to detect the presence of pulmonary embolism, 0.945 to detect acute pulmonary embolism, and 0.945 and 0.987 to detect central and subsegmental pulmonary embolisms, respectively (15).…”

Natural language processing to identify ureteric stones in radiology reports

“…Studies into NLP‐assisted cohort building for epidemiological research have shown promising results with sensitivities ranging from 85 to 98% and specificities ranging from 86 to 100% . Examples of applications include identifying reports containing renal cysts, pneumonia, pulmonary nodules, pulmonary embolism, metastases, adrenal nodules, peripheral arterial disease and misplaced lines or devices . To our knowledge there have been no published studies on NLP‐assisted detection of ureteric stones.…”