MedClassNet: a novel NLP framework for automated disease classification from clinical text using transformer and graph attention networks
Abstract
A combination of complex vocabulary, acronyms, and relationships in clinical narratives makes automatic disease classification from clinical narratives challenging. The study proposes the MedClassNet framework that combines Graph Attention Network (GAT) and BioBERT-based contextual embeddings for disease classification in clinical narratives. The MedClassNet framework uses BioBERT embeddings of admission and progress notes and generates a similarity graph for patients by using the cosine similarity between patient contextual embeddings. To classify diseases, the BioBERT framework is combined with the relational embedding learned by the GAT framework for clinically similar patients. An annotated corpus of 1,748 clinical narratives from MIMIC-III dataset containing single-label ICD-9 disease categories is used for experimentation. To avoid label leakage, discharge summaries, final diagnosis, explicit ICD-9 codes, and other post-diagnosis data are excluded from the input for model training. According to the experimental protocol, MedClassNet framework yields an accuracy of 94.2%, a precision of 94.6%, and F1-score of 0.93. Clinically relevant textual features responsible for predictions of the model are discovered by using SHAP-based analysis. The results indicate that combining patient-level relational information with clinical language contextual embeddings can improve disease classification performance as compared with baseline models.
Related
- An attention-based CNN–Transformer hybrid framework for osteoporosis and osteopenia classification in knee X-ray images
- Attention-integrated hybrid CNN–vision transformer framework for robust multi-class cervical cytology classification using liquid-based cytology images
- MangoSwin-attention: a swin transformer with attention pooling for generalizable mango leaf disease classification
- HM-Swin: Hierarchical Multiscale Swin Transformer Network with Attention Fusion for Heart Sound Classification
- SkinFusionNet: MobileNetV₂ and vision transformer hybrid deep learning model with automated skin cancer classification
BibTeX
@article{shivahare2026medclassnet,
title = {{MedClassNet: a novel NLP framework for automated disease classification from clinical text using transformer and graph attention networks}},
author = {Basu Dev Shivahare and Anjali Kapoor and Anju Mishra and Preeti Gupta and Sandeep Kumar Mathivanan and S. K. B. Sangeetha},
journal = {Scientific Reports},
year = {2026},
month = nov,
doi = {10.1038/s41598-026-76483-z},
url = {https://www.nature.com/articles/s41598-026-76483-z},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - MedClassNet: a novel NLP framework for automated disease classification from clinical text using transformer and graph attention networks
AU - Basu Dev Shivahare
AU - Anjali Kapoor
AU - Anju Mishra
AU - Preeti Gupta
AU - Sandeep Kumar Mathivanan
AU - S. K. B. Sangeetha
PY - 2026
DA - 2026/11/03
JO - Scientific Reports
DO - 10.1038/s41598-026-76483-z
UR - https://www.nature.com/articles/s41598-026-76483-z
AB - A combination of complex vocabulary, acronyms, and relationships in clinical narratives makes automatic disease classification from clinical narratives challenging. The study proposes the MedClassNet framework that combines Graph Attention Network (GAT) and BioBERT-based contextual embeddings for disease classification in clinical narratives. The MedClassNet framework uses BioBERT embeddings of admission and progress notes and generates a similarity graph for patients by using the cosine similarity between patient contextual embeddings. To classify diseases, the BioBERT framework is combined with the relational embedding learned by the GAT framework for clinically similar patients. An annotated corpus of 1,748 clinical narratives from MIMIC-III dataset containing single-label ICD-9 disease categories is used for experimentation. To avoid label leakage, discharge summaries, final diagnosis, explicit ICD-9 codes, and other post-diagnosis data are excluded from the input for model training. According to the experimental protocol, MedClassNet framework yields an accuracy of 94.2%, a precision of 94.6%, and F1-score of 0.93. Clinically relevant textual features responsible for predictions of the model are discovered by using SHAP-based analysis. The results indicate that combining patient-level relational information with clinical language contextual embeddings can improve disease classification performance as compared with baseline models.
SN - 2045-2322
ER -