Natural language inference for curation of structured clinical registries from unstructured text

Area of Science:

• Computational linguistics
• Medical informatics
• Oncology research

Background:

• Clinical registries are crucial for research and patient care but are costly and time-consuming to curate.
• Manual curation of registry data presents significant challenges, especially for rare diseases.

Purpose of the Study:

• To evaluate the feasibility of using natural language inference (NLI) models for scalable clinical registry curation.
• To assess the performance of state-of-the-art NLI models in extracting information for breast oncology registry fields.

Main Methods:

• Applied five deep learning-based NLI models (ALBERT, BART, RoBERTa, XLNet, ELECTRA) to clinical, laboratory, and pathology notes.
• Inferred information for 43 breast oncology registry fields and compared model inferences against a manually curated database of 7439 patients.

Main Results:

• NLI models exhibited varied performance across different registry fields.
• The ALBERT model outperformed others on 22 out of 43 fields.
• Errors were often due to misinterpretation of historical findings, abbreviations, and clinical term variations.

Conclusions:

• Pretrained NLI models can potentially curate multiple registry fields simultaneously without task-specific training.
• NLI represents an unexplored, efficient approach for clinical registry curation, enhancing data extraction from unstructured clinical text.