Symptom Recognition in Medical Conversations Via multi- Instance Learning and Prompt.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Electronic health record (EHR) systems require automatic symptom extraction from medical dialogues.
• Challenges include scattered clues, informal language, and ambiguous/negated statements in patient descriptions.
• Existing models struggle with unstructured, multi-turn conversations.

Purpose of the Study:

• To develop a novel approach for fine-grained symptom identification in medical dialogues.
• To improve the accuracy and efficiency of automated medical record generation.
• To address challenges of scattered information, non-standard terminology, and negation in patient-reported symptoms.

Main Methods:

• A multi-instance learning (MIL) framework treats conversations as bags of utterances for improved recall.
• Prompt-guided attention leverages standardized terminology to recognize synonyms, implicit mentions, and negations, enhancing precision.
• R-Drop regularization is used to improve robustness against noisy data.

Main Results:

• The proposed method achieved an 85.93% F1-score, outperforming baselines by approximately 8%.
• Achieved 85.09% precision and 86.83% recall, demonstrating strong performance.
• Successfully identified specific symptom-utterance pairs, showcasing fine-grained extraction capabilities.

Conclusions:

• The combined MIL and prompt-guided attention approach effectively addresses challenges in symptom extraction from medical dialogues.
• This method enables precise symptom documentation, advancing automated medical information extraction and intelligent EHR systems.
• The fine-grained extraction supports diagnostic decision support and improves the quality of medical records.