Comparative Analysis of Machine Learning Algorithms for Computer-Assisted Reporting Based on Fully Automated Cross-Lingual RadLex® Mappings

Objectives: Studies evaluating machine learning (ML) algorithms on cross-lingual RadLex® mappings for developing context-sensitive radiological reporting tools are lacking. Therefore, we investigated whether ML-based approaches can be utilized to assist radiologists in providing key imaging biomarkers – such as The Alberta stroke programme early CT score (APECTS). Material and Methods: A stratified random sample (age, gender, year) of CT reports (n=206) with suspected ischemic stroke was generated out of 3997 reports signed off between 2015-2019. Three independent, blinded readers assessed these reports and manually annotated clinico-radiologically relevant key features. The primary outcome was whether ASPECTS should have been provided (yes/no: 154/52). For all reports, both the findings and impressions underwent cross-lingual (German to English) RadLex®-mappings using natural language processing. Well-established ML-algorithms including classification trees, random forests, elastic net, support vector machines (SVMs) and boosted trees were evaluated in a 5 x 5-fold nested cross-validation framework. Further, a linear classifier (fastText) was directly fitted on the German reports. Ensemble learning was used to provide robust importance rankings of these ML-algorithms. Performance was evaluated using derivates of the confusion matrix and metrics of calibration including AUC, brier score and log loss as well as visually by calibration plots. Results: On this imbalanced classification task SVMs showed the highest accuracies both on human-extracted- (87%) and fully automated RadLex® features (findings: 82.5%; impressions: 85.4%). FastText without pre-trained language model showed the highest accuracy (89.3%) and AUC (92%) on the impressions. Ensemble learner revealed that boosted trees, fastText and SVMs are the most important ML-classifiers. Boosted trees fitted on the findings showed the best overall calibration curve. Conclusions: Contextual ML-based assistance suggesting ASPECTS while reporting neuroradiological emergencies is feasible, even if ML-models are restricted to be developed on limited and highly imbalanced data sets.