Automatic classification of scanned electronic health record documents

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Natural Language Processing

Background:

• Electronic Health Records (EHRs) often contain diverse scanned documents.
• Categorizing these documents is crucial for efficient clinical review.
• Unclassified documents can lead to increased clinician workload and potential information gaps.

Purpose of the Study:

• To design and evaluate a system for categorizing scanned documents within EHRs.
• To differentiate between clinically relevant and non-clinically relevant documents.
• To demonstrate the accuracy of text classification systems in this domain.

Main Methods:

• Text extraction from scanned documents using Optical Character Recognition (OCR).
• Development and evaluation of machine learning models, including deep learning (ClinicalBERT) and bag-of-words approaches.
• System evaluation across three classification levels (document, intermediate, individual classes) and various text processing methods.

Main Results:

• A deep learning model (ClinicalBERT) achieved the highest performance.
• Accuracies reached 0.973 for clinically relevant vs. non-clinically relevant classification.
• High accuracies were also obtained for intermediate (0.949) and individual class (0.913) distinctions.

Conclusions:

• Machine learning applied to OCR-extracted text can accurately classify scanned EHR content.
• Automated classification reduces the burden on clinicians by filtering relevant information.
• This technology has the potential to improve efficiency and reduce errors in clinical workflows.