R-GAT: cancer document classification leveraging graph-based residual network for scenarios with limited data

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Cancer Informatics
• Machine Learning for Healthcare

Background:

• Accurate classification of cancer-related biomedical abstracts is vital for cancer informatics and healthcare research.
• Limited labeled data and high computational costs of transformer models hinder progress in this field.

Purpose of the Study:

• To develop a computationally efficient and effective model for classifying cancer-related biomedical abstracts.
• To address the limitations of existing methods in terms of data requirements and computational resources.

Main Methods:

• Proposed a Residual Graph Attention Network (R-GAT) integrating multi-head attention and residual connections.
• Captured semantic and relational dependencies within biomedical texts.
• Evaluated the model on a curated dataset of 1,875 PubMed abstracts covering thyroid, colon, lung, and generic cancer topics.

Main Results:

• R-GAT achieved a macro-F1 score of 0.96 ± 0.01, demonstrating stable and competitive performance.
• Performance was comparable to transformer-based models (BioBERT, BioClinicalBERT) and classical baselines (Logistic Regression).
• R-GAT required significantly fewer computational resources than transformer models.

Conclusions:

• Lightweight graph-based architectures like R-GAT are reliable and resource-efficient alternatives to computationally intensive transformers in biomedical NLP.
• Attention mechanisms and residual connections are crucial for model robustness, especially with limited data.
• The curated dataset is released to promote reproducibility and further research in cancer informatics.