Development of message passing-based graph convolutional networks for classifying cancer pathology reports

Area of Science:

• Natural Language Processing
• Machine Learning
• Bioinformatics

Background:

• Graph convolutional networks (GCNs) like TextGCN are effective for natural language text classification.
• Existing TextGCN models face challenges with memory consumption and distribution.
• A novel Fast Message Passing Network (FastMPN) is proposed to address these limitations.

Purpose of the Study:

• To develop a more efficient and flexible GCN model for text classification.
• To apply the FastMPN model to clinical information extraction from cancer pathology reports.
• To extract key tumor properties including site, subsite, laterality, histology, behavior, and grade.

Main Methods:

• Implemented a GCN with a message passing architecture (FastMPN).
• Incorporated trainable node embedding and edge weights for enhanced model flexibility.
• Applied the FastMPN to a dataset of cancer pathology reports for information extraction.

Main Results:

• The FastMPN model demonstrated performance equivalent to or better than the multi-task convolutional neural network (MT-CNN) model.
• Evaluated using micro- and macro-averaged F1 scores for clinical task performance.
• Achieved highly competitive clinical task performance scores for tumor information extraction.

Conclusions:

• The FastMPN model, implemented on PyTorch, efficiently trains large corpora (667,290 samples) in under 3 minutes per epoch on a V100 GPU.
• The FastMPN offers a versatile and flexible approach to GCN-based text classification.
• This implementation provides a competitive solution for extracting critical tumor information from pathology reports.