Leveraging Summary Guidance on Medical Report Summarization

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Automated summarization of medical texts is crucial for efficient information retrieval.
• Existing datasets and methods may not fully capture the nuances of clinical documentation.
• Large-scale, deidentified medical datasets are needed for robust model training.

Purpose of the Study:

• To introduce three novel, large-scale, deidentified medical text datasets: DISCHARGE, ECHO, and RADIOLOGY.
• To establish baseline performance for automated abstractive summarization using pre-trained language models.
• To propose and evaluate an enhanced summarization method leveraging prior knowledge guidance.

Main Methods:

• Creation of DISCHARGE (50K), ECHO (16K), and RADIOLOGY (378K) report-summary pairs from MIMIC-III.
• Implementation and evaluation of multiple pre-trained encoder-decoder models (BERT2BERT, BART, T5, etc.).
• Development of a BART-based model incorporating sampled training summaries as prior knowledge for improved encoding and decoding.

Main Results:

• Established baseline performance metrics for automated summarization on the new datasets.
• Demonstrated significant improvements in ROUGE scores and BERTScore using the proposed guided summarization method.
• Validated the effectiveness of incorporating prior knowledge into the summarization process.

Conclusions:

• The introduced datasets facilitate advancements in medical text summarization research.
• The proposed guided summarization approach enhances the quality of automated clinical note summaries.
• This work contributes to more effective clinical information management through improved NLP techniques.