SSMT-PANBERT: A single-stage multitask model for phenotype extraction and assertion negation detection in...

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Artificial Intelligence in Healthcare

Background:

• Electronic Health Records (EHRs) contain unstructured clinical notes vital for healthcare applications.
• Manual processing of EHRs is challenging due to data volume and complexity.
• Existing automated methods for phenotype extraction and negation detection have limitations in handling complex clinical scenarios.

Purpose of the Study:

• To develop a single-stage multitask solution for joint phenotype extraction and assertion negation detection from EHRs.
• To leverage state-of-the-art pre-trained language models (PLMs) for improved accuracy and efficiency.
• To provide practical assistance to healthcare professionals by handling diverse clinical data.

Main Methods:

• A single-stage multitask learning framework was proposed, integrating phenotype extraction and assertion negation detection.
• The approach utilizes advanced pre-trained language models (PLMs) for end-to-end processing.
• The method was evaluated on a validated dataset derived from MIMIC-III clinical notes, with annotations reviewed by domain experts.

Main Results:

• The top-performing model, SSMT-PANBERT, achieved high performance with an average Macro F1 score of 92.33% and Micro F1 score of 91.66%.
• The unified approach outperformed traditional pipeline methods in Macro F1 score.
• Significant computational advantages were observed, including a 37% reduction in training time, 18.2% in inference time, and 57% in GPU memory usage.

Conclusions:

• The proposed unified approach effectively handles complex clinical scenarios for phenotype extraction and negation detection.
• The method offers substantial computational benefits, making it suitable for real-world healthcare applications.
• Further analysis identified areas for future model improvement and refinement.