Improving the Factual Accuracy of Abstractive Clinical Text Summarization using Multi-Objective Optimization.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Artificial Intelligence

Background:

• Abstractive summarization techniques show promise across various domains but have seen limited application in clinical text.
• Clinical notes present unique challenges due to their unstructured nature and the scarcity of large-scale training data.
• Ensuring factual accuracy in clinical text summaries is paramount, especially in critical fields like cardiology.

Purpose of the Study:

• To propose a framework for enhancing the factual accuracy of abstractive summarization for clinical text.
• To address the limitations of current clinical text summarization methods, particularly concerning data availability and note structure.
• To improve the reliability of generated clinical summaries for better patient care.

Main Methods:

• Developed a knowledge-guided multi-objective optimization framework for abstractive summarization.
• Proposed joint optimization of three cost functions: generative loss, entity loss, and knowledge loss.
• Evaluated the framework on a newly collected dataset of heart failure (HF) clinical notes and two benchmark datasets (IU X-Ray, MIMIC-CXR).

Main Results:

• The proposed framework demonstrated improved performance in entity-level factual accuracy for clinical text summarization.
• Experiments were conducted using three distinct transformer encoder-decoder architectures.
• Optimizing the defined loss functions led to significant gains in the factual precision of generated summaries.

Conclusions:

• The knowledge-guided multi-objective optimization framework effectively enhances factual accuracy in clinical text summarization.
• This approach holds significant potential for improving the quality and reliability of clinical summaries in cardiology and beyond.
• Further research can explore broader applications and refinements of this technique in healthcare AI.