Extraction of radiographic findings from unstructured thoracoabdominal computed tomography reports using convolutional neural network based natural language processing.

BACKGROUND

Heart failure (HF) is a major cause of morbidity and mortality. However, much of the clinical data is unstructured in the form of radiology reports, while the process of data collection and curation is arduous and time-consuming.

PURPOSE

We utilized a machine learning (ML)-based natural language processing (NLP) approach to extract clinical terms from unstructured radiology reports. Additionally, we investigate the prognostic value of the extracted data in predicting all-cause mortality (ACM) in HF patients.

MATERIALS AND METHODS

This observational cohort study utilized 122,025 thoracoabdominal computed tomography (CT) reports from 11,808 HF patients obtained between 2008 and 2018. 1,560 CT reports were manually annotated for the presence or absence of 14 radiographic findings, in addition to age and gender. Thereafter, a Convolutional Neural Network (CNN) was trained, validated and tested to determine the presence or absence of these features. Further, the ability of CNN to predict ACM was evaluated using Cox regression analysis on the extracted features.

RESULTS

11,808 CT reports were analyzed from 11,808 patients (mean age 72.8 ± 14.8 years; 52.7% (6,217/11,808) male) from whom 3,107 died during the 10.6-year follow-up. The CNN demonstrated excellent accuracy for retrieval of the 14 radiographic findings with area-under-the-curve (AUC) ranging between 0.83-1.00 (F1 score 0.84-0.97). Cox model showed the time-dependent AUC for predicting ACM was 0.747 (95% confidence interval [CI] of 0.704-0.790) at 30 days.

CONCLUSION

An ML-based NLP approach to unstructured CT reports demonstrates excellent accuracy for the extraction of predetermined radiographic findings, and provides prognostic value in HF patients.

Déclaration de conflit d'intérêts

The authors have declared that no competing interests exist. Gurpreet Singh is currently employed at GlaxoSmithKline but was not a part of GlaxoSmithKline during the conduct of this study. Gabriel Maliakal and James K. Min are currently employed at Cleerly Inc. but were not a part of Cleerly Inc. during the conduct of this study. Mohit Pandey is currently employed at Ipsos but was not a part of Ipsos during the conduct of this study. These commercial affiliations do not alter our adherence to PLOS ONE policies on sharing data and materials.