Closing the loop: automatically identifying abnormal imaging results in scanned documents.

Area of Science:

• Medical informatics
• Natural Language Processing in Healthcare
• Radiology reporting

Background:

• Scanned documents (SDs) in electronic health records contain valuable clinical information but are difficult to integrate into clinician workflows.
• Identifying critical findings in imaging reports, such as potential malignancies or fractures, is crucial for timely patient follow-up and malpractice claim prevention.

Purpose of the Study:

• To develop and evaluate a ClinicalBERT-based pipeline for automatically identifying clinically significant abnormalities in scanned imaging reports.
• To improve the efficiency of reviewing scanned reports by reducing manual screening efforts.

Main Methods:

• A ClinicalBERT-based natural language processing pipeline was trained on manually classified and ICD-10 coded scanned reports.
• The pipeline was evaluated on a test set of manually classified scanned imaging reports (mammograms, chest CTs, bone X-rays).
• Performance was compared against a baseline string-matching approach, optimizing for both F1 score and recall.

Main Results:

• The developed pipeline significantly outperformed string-matching, achieving F1 scores up to 0.905 for mammograms and 0.900 for chest CTs.
• Models optimized for recall achieved a 1.000 recall for all report types, with precisions ranging from 0.275 to 0.727.
• Models trained on ICD-10 codes also showed competitive performance, with F1 scores up to 0.830 for chest CTs.

Conclusions:

• Automated identification of clinically significant abnormalities in scanned imaging reports is feasible with high recall and practical precision.
• This pipeline can reduce the manual effort required for screening abnormal findings, enhancing clinician workflow and patient care.
• The approach is generalizable and minimally laborious, offering a practical solution for managing unstructured data in electronic health records.