Diabetes Prediction with Code-Based Mecical Insurance Claims Based on Multimodal Representations and Vision-Language...

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Public Health

Background:

• Rising prevalence of lifestyle-related chronic diseases like diabetes and hyperlipidemia strains healthcare systems globally.
• Accurate forecasting of future medical demand is essential for effective resource allocation and planning.
• Public medical insurance claims (MICs) offer a valuable, large-scale data source for health trend analysis.

Purpose of the Study:

• To develop a predictive model for future healthcare demand using Japanese medical insurance claims (MICs).
• To enhance the accuracy of demand prediction by leveraging large language models (LLMs) for semi-structured, code-based insurance data.
• To address the limitations of traditional LLM applications on complex medical claim data.

Main Methods:

• Proposed a multimodal representation technique tailored for code-based medical insurance claims (MICs).
• Developed and applied a vision-language model (VLM) for predicting future medical demand.
• Utilized a large dataset of public medical insurance claims from Japan's universal healthcare system.

Main Results:

• The proposed multimodal VLM significantly improved prediction accuracy compared to baseline methods.
• Achieved a 3.8-point increase in accuracy for predicting diabetes-related cases.
• Demonstrated the effectiveness of the multimodal approach for complex, semi-structured medical data.

Conclusions:

• The multimodal VLM approach offers a promising solution for accurate, data-driven healthcare demand prediction.
• This method can aid policymakers and healthcare providers in proactive resource management and planning.
• Exploiting LLMs with specialized data representations enhances predictive capabilities for public health challenges.