CATI: A medical context-enhanced framework for diagnosis code assignment in the UK Biobank study

Area of Science:

• Bioinformatics
• Medical Informatics
• Computational Biology

Background:

• Accurate diagnosis coding is essential for large-scale biobank studies and downstream disease research.
• Current methods often neglect valuable textual medical information and the hierarchical nature of disease classification systems.
• Missing diagnosis codes in biobanks limit the utility of patient data for clinical research.

Purpose of the Study:

• To develop and evaluate CATI, a medical context-enhanced framework for improved diagnosis code assignment in large biobanks.
• To integrate textual details and disease hierarchy into a unified model for more accurate coding.
• To address the challenge of missing diagnosis codes for patients in biobank datasets.

Main Methods:

• CATI framework integrates textual details from key features and disease hierarchy using text embeddings from BioBERT via prompt tuning.
• A novel convolution layer is developed to effectively propagate information between adjacent diagnosis codes, leveraging the hierarchical structure.
• The study utilizes UK Biobank data, focusing on Phecodes and ICD-10 codes as standard disease formats.

Main Results:

• CATI demonstrates superior performance compared to state-of-the-art methods for both Phecodes and ICD-10 code assignment.
• Achieved at least a 5.16% improvement in average AUROC for unseen disease codes.
• Showcased an 8.68% increase in average AUPRC for disease codes with limited training instances (1000-10000).

Conclusions:

• CATI effectively enhances diagnosis code assignment by incorporating medical context from textual data and disease hierarchies.
• The framework facilitates the creation of well-defined cohorts for downstream biomedical research.
• CATI offers a valuable approach for complex healthcare tasks by leveraging rich medical information.