Recurrent neural networks for classifying relations in clinical notes.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Machine Learning

Background:

• Accurate identification of relationships between medical concepts in clinical notes is crucial for healthcare.
• Previous methods often rely on extensive manual feature engineering, which is time-consuming and labor-intensive.

Purpose of the Study:

• To introduce the first recurrent neural network (RNN) based models, specifically Long Short-Term Memory (LSTM), for clinical relation classification.
• To evaluate the performance of these models on a standard benchmark dataset and compare them against existing state-of-the-art systems.

Main Methods:

• Development and application of segment LSTM and sentence LSTM models for relation classification.
• Utilizing word embeddings, including domain-specific medical embeddings, as features.
• Comparison of model performance using micro-averaged f-measure across different relation types.

Main Results:

• The segment LSTM model achieved competitive micro-averaged f-measures: 0.661 (problem-treatment), 0.800 (problem-test), and 0.683 (problem-problem).
• The segment LSTM model outperformed the sentence LSTM model, highlighting the importance of contextual text analysis.
• Medical domain word embeddings demonstrated an improvement in LSTM model performance.

Conclusions:

• LSTM models are effective for classifying relations between medical concepts in clinical text.
• These models offer comparable performance to state-of-the-art systems while eliminating the need for manual feature engineering.
• The findings support the adoption of LSTM networks for automated clinical relation extraction.