Key Verbatim Extraction from Clinical Notes: A Hierarchical Multimodal Cross-Attention Approach
Abstract: Clinical notes are essential for physicians to accurately assess patient conditions, particularly in oncology where records are extensive. Efficient and effective information extraction from these notes is crucial for effective treatment. This is not a trivial task due to the lengthy and specialized content in the notes. Current methods that capture token-level or sentence-level relations, which are context-dependent, are sometimes insufficient for knowledge-intensive tasks such as information extraction from EHR that require external knowledge. To address this, we introduce a knowledge-enhanced hierarchical multimodal cross-attention approach. This method employs a cross-attention mechanism to integrate textual knowledge with patient network knowledge, aiming to synthesize information across multiple data levels, including word, sentence, note, and patient levels. This approach can efficiently highlight key sentences in clinical notes. We validate our method using extensive experiments on a large real-world dataset. The results demonstrate that our proposed model outperforms baseline models by up to 4.17% and 2.79% regarding F1 and accuracy.
Paper Type: Long
Research Area: Information Extraction
Research Area Keywords: healthcare applications, multimodality, clinical NLP, cross-modal information extraction
Languages Studied: English
Submission Number: 5507
Loading