Family History Information Extraction With Neural Attention and an Enhanced Relation-Side Scheme: Algorithm...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Family history information (FHI) is crucial for assessing disease susceptibility but is often embedded in narrative clinical text.
• Extracting FHI automatically from electronic health records (EHRs) requires advanced natural language processing (NLP) techniques.
• Current methods face challenges in inferring family member details and leveraging information across sentences.

Purpose of the Study:

• To develop an NLP approach for accurate family history information extraction from clinical reports.
• To address limitations in previous FHI extraction methods, specifically the need for postprocessing rules and efficient use of contextual information.
• To improve the utilization of both intra- and inter-sentence information for enhanced FHI extraction.

Main Methods:

• Formulated FHI extraction as a sequential labeling problem.
• Proposed an enhanced relation-side scheme to encode family member properties, eliminating the need for postprocessing rules.
• Developed an attention-based neural network to effectively utilize cross-sentence information for FHI attribute identification.

Main Results:

• The enhanced scheme improved neural network recall, increasing the F score by 0.024.
• The attention-based neural network enhanced both recall and precision, achieving an F score of 0.807.
• The proposed method achieved a fourth-place ranking in the 2019 n2c2/OHNLP family history extraction task.

Conclusions:

• An attention-based neural network combined with an enhanced tag scheme was presented.
• The model effectively learns and interprets implicit relationships and side information of family members across sentences.
• This approach eliminates the reliance on heuristic rules for accurate FHI extraction.