Weakly supervised spatial relation extraction from radiology reports.

Area of Science:

• Clinical Natural Language Processing
• Medical Informatics
• Artificial Intelligence in Radiology

Background:

• Manual annotation of clinical data is time-consuming and expensive.
• Weak supervision offers a promising alternative for training NLP models using existing resources.
• Extracting spatial information from radiology reports is crucial for clinical decision-making.

Purpose of the Study:

• To evaluate a weak supervision approach for extracting spatial information from radiology reports.
• To assess the performance of a data programming-based weak supervision method.
• To compare the efficacy of weakly supervised models with and without manual fine-tuning.

Main Methods:

• Developed a weak supervision framework using data programming and labeling functions based on domain-specific dictionaries.
• Generated weak labels for spatial relations in radiology reports.
• Fine-tuned a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model using these weak labels.

Main Results:

• The weakly supervised BERT model achieved an F1 score of 72.89 for spatial trigger extraction and 52.47 for relation extraction without manual annotations.
• Fine-tuning the model on manual annotations improved relation extraction performance to an F1 score of 68.76.
• The model's performance surpassed fully supervised state-of-the-art results after fine-tuning.

Conclusions:

• Weak supervision provides a viable and efficient method for extracting critical spatial information from radiology reports.
• The proposed data programming approach is adaptable and generalizable across different radiology subdomains.
• This method demonstrates the potential to significantly advance clinical NLP by reducing reliance on extensive manual annotation.