Transformer-Enhanced Joint Information Extraction Framework for Bladder Cancer Medical Dialogue Analysis

Area of Science:

• Medical Informatics
• Natural Language Processing
• Oncology

Background:

• Extracting clinical information from medical dialogues for bladder cancer diagnosis and treatment monitoring is challenging for single-task approaches.
• Patient-physician conversations contain scattered, subtle indicators of symptoms, treatment responses, and disease progression.

Purpose of the Study:

• To develop a novel framework, bladder cancer transformer for joint information extraction (BCT-JIE), for simultaneous information extraction from bladder cancer medical dialogues.
• To improve the accuracy of entity recognition, intent classification, and clinical state assessment in bladder cancer conversations.

Main Methods:

• A collaborative transformer architecture (BCT-JIE) integrating sliding window mechanisms, bidirectional long short-term memory networks, and a symptom-aware module.
• Global attention mechanisms to capture long-range dependencies and interactive guidance modules to model task interdependencies.

Main Results:

• BCT-JIE achieved superior performance over baseline methods on MedDialog and Multi MedQA datasets.
• Demonstrated significant improvements in bladder cancer-specific entity extraction (91.76% F1-score) and clinical state classification (83.42% accuracy).
• Ablation studies confirmed the contributions of the sliding window and global attention mechanisms.

Conclusions:

• The BCT-JIE framework effectively addresses the complexities of information extraction in bladder cancer medical dialogues.
• This approach enhances the accuracy of clinical information capture, supporting improved bladder cancer management and monitoring.