Graph augmented transformers improve chemotherapy toxicity symptom extraction from clinical notes

Chemotherapy is essential for cancer treatment but may cause adverse events requiring emergency department visits and hospitalizations, placing substantial burdens on patients and healthcare systems. Existing approaches to detect these events often rely on structured electronic health records (EHR) data, which incompletely capture patients' symptom trajectories. Clinical notes contain richer information yet remain challenging to synthesize. Here we show that integrating transformer-based language models with graph neural networks improves extraction of chemotherapy-related toxicity symptoms from clinical notes. We developed Graph-Augmented Transformer for Clinical Notes (GAT-CN), which embeds patient notes using Bio+ClinicalBERT and links them to symptom-related terms within a heterogeneous clinical graph learned using GraphSAGE. In a multi-symptom classification task, GAT-CN outperformed transformer-only models, achieving a weighted AUROC of 0.850 and AUPRC of 0.812. The model also identified additional diagnoses not captured in structured EHRs, confirmed through manual annotation. These results demonstrate that graph-augmented models improve symptom detection from clinical narratives and support earlier monitoring of chemotherapy-related adverse events.