Toward structuring real-world data: Deep learning for extracting oncology information from clinical text with...

Area of Science:

• Computational biology
• Medical informatics
• Natural Language Processing

Background:

• Detailed patient information is primarily in unstructured clinical notes within real-world data (RWD).
• Manual data curation is costly and inefficient, hindering real-world evidence (RWE) generation.
• Structuring RWD is crucial for advancing RWE studies.

Purpose of the Study:

• To develop and evaluate advanced deep-learning natural language processing (NLP) methods for structuring RWD.
• To utilize patient-level supervision from medical registries for general RWD applications.
• To assess the performance of NLP methods on a large-scale cancer registry dataset.

Main Methods:

• Leveraged patient-level supervision from a cancer registry within a large integrated delivery network (IDN).
• Employed deep-learning models trained on 135,107 patients across five western US states.
• Validated performance on held-out data from separate health systems and states.

Main Results:

• Achieved high test area under the receiver operating characteristic curve (AUROC) values of 94%-99% for key tumor attributes.
• Demonstrated comparable performance on data from different health systems and states.
• Ablation studies confirmed the superiority of the proposed deep-learning methods.
• Identified instances where the NLP system corrected errors in the original registrar labels.

Conclusions:

• Deep-learning NLP methods effectively structure real-world data from clinical notes using medical registry supervision.
• The developed system shows high accuracy and generalizability across different healthcare settings.
• This approach offers a scalable solution for real-world evidence generation and can potentially improve data quality.