A Disease-Prediction Protocol Integrating Triage Priority and BERT-Based Transfer Learning for Intelligent Triage

Area of Science:

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

Background:

• Large hospitals present complex navigation challenges for patients with limited medical knowledge.
• Misdirected patient visits and unnecessary appointments are common, leading to inefficiencies.
• There is a need for intelligent, remote self-service triage systems in modern healthcare.

Purpose of the Study:

• To develop an intelligent triage system for processing multilabel neurological medical texts.
• To predict patient diagnoses and appropriate hospital departments using transfer learning.
• To enhance patient self-service triage capabilities and improve hospital workflow.

Main Methods:

• Utilized transfer learning and the BERT model for classifying chief complaint text.
• Developed a triage priority (TP) method to convert multilabel classification into single-label.
• Incorporated a composite loss function with cost-sensitive learning to address data imbalance and disease severity.

Main Results:

• The triage priority (TP) method achieved 87.47% accuracy in classifying medical record text.
• The enhanced BERT model with a composite loss function improved accuracy to 88.38%.
• The system demonstrated improved triage accuracy and reduced patient confusion compared to traditional methods.

Conclusions:

• The intelligent triage system effectively processes neurological medical texts for accurate diagnosis and department prediction.
• The proposed TP method and composite loss function offer significant improvements in classification accuracy.
• This system enhances hospital triage capabilities, streamlines patient flow, and improves the overall patient experience.