Temporal bone radiology report classification using open source machine learning and natural langue processing libraries
Related Experiment Video
Updated: Mar 19, 2026
High-Speed Human Temporal Bone Sectioning for the Assessment of COVID-19-Associated Middle Ear Pathology
Published on: May 18, 2022
Aaron J Masino1, Robert W Grundmeier2,3, Jeffrey W Pennington2
1Department of Biomedical and Health Informatics, The Children's Hospital of Philadelphia, 3535 Market Street, Suite 1024, Philadelphia, PA, 19104, USA. masinoa@email.chop.edu.
Background:
Radiology reports are a rich resource for biomedical research. Prior to utilization, trained experts must manually review reports to identify discrete outcomes. The Audiological and Genetic Database (AudGenDB) is a public, de-identified research database that contains over 16,000 radiology reports. Because the reports are unlabeled, it is difficult to select those with specific abnormalities. We implemented a classification pipeline using a human-in-the-loop machine learning approach and open source libraries to label the reports with one or more of four abnormality region labels: inner, middle, outer, and mastoid, indicating the presence of an abnormality in the specified ear region.
Methods:
Trained abstractors labeled radiology reports taken from AudGenDB to form a gold standard. These were split into training (80 %) and test (20 %) sets. We applied open source libraries to normalize and convert every report to an n-gram feature vector. We trained logistic regression, support vector machine (linear and Gaussian), decision tree, random forest, and naïve Bayes models for each ear region. The models were evaluated on the hold-out test set.
Results:
Our gold-standard data set contained 726 reports. The best classifiers were linear support vector machine for inner and outer ear, logistic regression for middle ear, and decision tree for mastoid. Classifier test set accuracy was 90 %, 90 %, 93 %, and 82 % for the inner, middle, outer and mastoid regions, respectively. The logistic regression method was very consistent, achieving accuracy scores within 2.75 % of the best classifier across regions and a receiver operator characteristic area under the curve of 0.92 or greater across all regions.
Conclusions:
Our results indicate that the applied methods achieve accuracy scores sufficient to support our objective of extracting discrete features from radiology reports to enhance cohort identification in AudGenDB. The models described here are available in several free, open source libraries that make them more accessible and simplify their utilization as demonstrated in this work. We additionally implemented the models as a web service that accepts radiology report text in an HTTP request and provides the predicted region labels. This service has been used to label the reports in AudGenDB and is freely available.
