Automatic Disease Annotation From Radiology Reports Using Artificial Intelligence Implemented by a Recurrent Neural...

Area of Science:

• Medical Informatics
• Radiology
• Artificial Intelligence

Background:

• Radiology reports are valuable for research but require manual annotation, which is time-consuming.
• Automating the categorization of free-text electronic medical record (EMR) data, like radiology reports, presents significant challenges.
• Manual review of radiology reports for disease annotation is a bottleneck in biomedical research.

Purpose of the Study:

• To develop an automated system for disease annotation in radiology reports.
• To create a recurrent neural network (RNN)-based system for classifying fracture and non-fracture cases.
• To improve the efficiency of analyzing large volumes of radiology report data.

Main Methods:

• A recurrent neural network (RNN) model was developed to automatically identify fracture and non-fracture cases.
• The system was trained and tested on 3032 sentences from musculoskeletal radiography reports, manually classified by orthopedic surgeons.
• Performance was evaluated using word error rate and standard classification metrics (accuracy, precision, recall, F1 score).

Main Results:

• The three-layer RNN model achieved the best performance with a word error rate of 1.03% using Levenshtein distance.
• The model demonstrated high performance as a binary classifier, with precision, recall, and F1 score of 0.967, and accuracy of 0.982.
• The system effectively classified important findings in radiology reports.

Conclusions:

• The RNN-based system demonstrates a strong ability to classify key findings in radiology reports, evidenced by a high F1 score.
• This automated system offers an efficient solution for analyzing extensive radiology report data.
• The developed system has the potential for application in cohort construction for retrospective studies.