A Frame-Based NLP System for Cancer-Related Information Extraction

Area of Science:

• Medical Informatics
• Computational Linguistics
• Clinical Data Mining

Background:

• Clinical narratives contain valuable cancer-related information that is challenging to extract.
• Existing methods for information extraction from clinical text often lack the necessary depth and accuracy.
• Frame semantics offer a structured way to represent complex information within narratives.

Purpose of the Study:

• To develop and evaluate a frame-based natural language processing (NLP) method for extracting cancer-related information from clinical narratives.
• To focus on three key frames: cancer diagnosis, cancer therapeutic procedure, and tumor description.
• To assess the effectiveness of deep learning models, specifically bidirectional Long Short-term Memory (LSTM) Conditional Random Field (CRF), for this task.

Main Methods:

• Utilized a deep learning approach combining bidirectional Long Short-term Memory (LSTM) and Conditional Random Field (CRF).
• Employed both character and word embeddings, including GloVe and MIMIC-III, for enhanced representation.
• Developed a system with two sequence classifiers: frame identification (lexical unit) and frame element classification.

Main Results:

• Achieved high F1-scores: 93.70 for cancer diagnosis, 96.33 for therapeutic procedure, and 87.18 for tumor description.
• Demonstrated significant improvements over a baseline heuristic, with gains of 10.72, 0.85, and 8.04, respectively.
• Found that combining GloVe and MIMIC-III embeddings yielded the best representational performance.

Conclusions:

• The proposed frame-based NLP method effectively extracts frame semantic information from clinical narratives.
• Deep learning techniques, particularly the LSTM-CRF architecture, show strong performance in clinical information extraction.
• This approach offers a promising avenue for improving the utilization of unstructured clinical data for cancer research and patient care.