Graph augmented transformers improve chemotherapy toxicity symptom extraction from clinical notes.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Linguistics
• Oncology

Background:

• Chemotherapy is vital for cancer treatment but can cause adverse events, leading to significant healthcare costs.
• Current methods using structured electronic health records (EHR) fail to capture the full spectrum of patient symptoms.
• Clinical notes offer rich data but are difficult to analyze effectively.

Purpose of the Study:

• To develop and evaluate an advanced AI model for improved extraction of chemotherapy-related toxicity symptoms from clinical notes.
• To enhance the early detection and monitoring of adverse events associated with cancer chemotherapy.

Main Methods:

• Developed Graph-Augmented Transformer for Clinical Notes (GAT-CN), integrating Bio+ClinicalBERT embeddings with GraphSAGE-learned heterogeneous clinical graphs.
• Utilized transformer-based language models and graph neural networks to process and analyze unstructured clinical text.
• Performed multi-symptom classification to assess model performance against transformer-only baselines.

Main Results:

• GAT-CN achieved a weighted AUROC of 0.850 and AUPRC of 0.812 in multi-symptom classification, outperforming transformer-only models.
• The model successfully identified additional diagnoses missed in structured EHR data, validated by manual annotation.
• Demonstrated superior performance in extracting chemotherapy toxicity symptoms from narrative clinical data.

Conclusions:

• Graph-augmented models significantly enhance the detection of patient symptoms from clinical narratives.
• The GAT-CN model shows promise for earlier and more accurate monitoring of chemotherapy-related adverse events.
• Integrating language models with graph neural networks offers a powerful approach to analyzing complex clinical text for improved patient outcomes.