Memory-Augmented Active Deep Learning for Identifying Relations Between Distant Medical Concepts in...

Area of Science:

• Medical Informatics
• Computational Neuroscience
• Natural Language Processing

Background:

• Automatic identification of medical concept relations in Electroencephalography (EEG) reports is crucial for building EEG-specific patient retrieval systems and knowledge acquisition.
• Extracting EEG-specific relations is challenging due to long-distance dependencies, where concepts appear in different sentences or sections.

Purpose of the Study:

• To develop a novel framework for identifying long-distance relations between medical concepts in EEG reports.
• To improve the accuracy and efficiency of knowledge extraction from large EEG report corpora.

Main Methods:

• A novel framework combining Dynamic Relational Memory (DRM), a deep learning approach, with active learning was developed.
• DRM facilitates the prediction of long-distance relations, while active learning optimizes relation identification with minimal training data.

Main Results:

• The framework achieved a 5-fold cross-validation F1 score of 0.7475 on 140 EEG reports selected via active learning.
• The combined DRM and active learning approach demonstrated significant promise in handling long-distance dependencies for relation extraction.

Conclusions:

• The proposed framework effectively addresses the challenge of identifying long-distance relations in EEG reports.
• This approach shows great potential for advancing EEG-specific patient cohort retrieval and knowledge discovery.