Paper Type

Complete

Paper Number

PACIS2025-1545

Description

Recent advancements in artificial intelligence are revolutionizing healthcare. The development of AI models for predicting patient acuity in tele-triage shows promising potential for high-quality patient care and efficient healthcare systems. However, existing research has not explored how to achieve effective predictions using limited patient self-reported data. Therefore, we follow the computational design science approach to propose DAJoLoss, a novel dual-adapter architecture with a domain-based joint loss optimization method for fine-tuning large language models (LLMs). This computational artifact enables transfer learning from informative training-only data to improve prediction. Comprehensive evaluation results demonstrate that our artifact achieves at least a 5% improvement across all metrics compared to mainstream machine learning approaches in tele-triage research. This study not only provides insights into overcoming data availability limitations in tele-triage prediction but also contributes to information systems knowledge base on designing computational artifacts for optimizing LLMs to make methodological contributions toward addressing grand social challenges.

Comments

Healthcare

Share

COinS
 
Jul 6th, 12:00 AM

Predicting Patient Acuity in Tele-Triage Using Large Language Models: A Computational Design Science Approach

Recent advancements in artificial intelligence are revolutionizing healthcare. The development of AI models for predicting patient acuity in tele-triage shows promising potential for high-quality patient care and efficient healthcare systems. However, existing research has not explored how to achieve effective predictions using limited patient self-reported data. Therefore, we follow the computational design science approach to propose DAJoLoss, a novel dual-adapter architecture with a domain-based joint loss optimization method for fine-tuning large language models (LLMs). This computational artifact enables transfer learning from informative training-only data to improve prediction. Comprehensive evaluation results demonstrate that our artifact achieves at least a 5% improvement across all metrics compared to mainstream machine learning approaches in tele-triage research. This study not only provides insights into overcoming data availability limitations in tele-triage prediction but also contributes to information systems knowledge base on designing computational artifacts for optimizing LLMs to make methodological contributions toward addressing grand social challenges.