Identifying Hepatocellular Carcinoma from imaging reports using natural language processing to facilitate data extraction from electronic patient records

BackgroundThe National Institute for Health Research Health Informatics Collaborative (NIHR HIC) viral hepatitis theme is working to overcome governance and data challenges to collate routine clinical data from electronic patients records from multiple UK hospital sites for translational research. The development of hepatocellular carcinoma (HCC) is a critical outcome for patients with viral hepatitis with the drivers of cancer transformation poorly understood. ObjectiveThis study aims to develop a natural language processing (NLP) algorithm for automatic HCC identification from imaging reports to facilitate studies into HCC. Methods1140 imaging reports were retrieved from the NIHR HIC viral hepatitis research database v1.0. These reports were from two sites, one used for method development (site 1) and the other for validation (site 2). Reports were initially manually annotated as binary classes (HCC vs. non-HCC). We designed inference rules for recognising HCC presence, wherein medical terms for eligibility criteria of HCC were determined by domain experts. A rule-based NLP algorithm with five submodules (regular expressions of medical terms, terms recognition, negation detection, sentence tagging, and report label generation) was developed and iteratively tuned. ResultsOur rule-based algorithm achieves an accuracy of 99.85% (sensitivity: 90%, specificity: 100%) for identifying HCC on the development set and 99.59% (sensitivity: 100%, specificity: 99.58%) on the validation set. This method outperforms several off-the-shelf models on HCC identification including "machine learning based" and "deep learning based" text classifiers in achieving significantly higher sensitivity. ConclusionOur rule-based NLP method gives high sensitivity and high specificity for HCC identification, even from imbalanced datasets with a small number positive cases, and can be used to rapidly screen imaging reports, at large-scale to facilitate epidemiological and clinical studies into HCC. Statement of Significance ProblemEstablishing a cohort of hepatocellular carcinoma (HCC) from imaging reports via manual review requires advanced clinical knowledge and is costly, time consuming, impractical when performed on a large scale. What is Already KnownAlthough some studies have applied natural language processing (NLP) techniques to facilitate identifying HCC information from narrative medical data, the proposed methods based on a pre-selection by diagnosis codes, or subject to certain standard templates, have limitations in application. What This Paper AddsWe have developed a hierarchical rule-based NLP method for automatic identification of HCC that uses diagnostic concepts and tumour feature representations that suggest an HCC diagnosis to form reference rules, accounts for differing linguistic styles within reports, and embeds a data pre-processing module that can be configured and customised for different reporting formats. In doing so we have overcome major challenges including the analysis of imbalanced data (inherent in clinical records) and lack of existing unified reporting standards.