Detection of unexpected findings in radiology reports: A comparative study of machine learning approaches

Abstract

This study explores machine learning methods for the detection of unexpected findings in Spanish radiology reports. Regarding radiological reports, unexpected findings are the set of radiological signs identified at a certain imaging modality exam which meet two characteristics: they are not apparently related with the a priori expected results of the radiological exam and involve a clinical emergency or urgency situation that must be reported shortly to the prescribing physician or another medical specialist as well as to the patient in order to preserve life and/or prevent dangerous occurrences. Several traditional machine learning and deep learning classification algorithms are evaluated and compared. To carry out the task we use 5947 anonymous radiology reports from HT médica. Experimental results suggest that the performance of the Convolutional Neural Networks models are better than traditional machine learning. The best F1 score for the identification of an unexpected finding was 90%. Finally, we also perform an error analysis which will guide us to achieve better results in the future.