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

Chronic diseases associated with lifestyles, such as diabetes and hyperlipidemia, are growing in many countries. That leads to increased social security expenditures and insufficient medical resources. The optimization of data-driven medical systems is crucial for medical facilities and governments to make accurate medical planning and provide their citizens with quality medical services. A solution should be future demand prediction, allowing the authorities to understand the care demand that might be onset and to secure necessary medical resources in advance. This work aims to enable the prediction of future demands from public medical insurance claims (MICs). Japan has implemented a universal medical care insurance system in which citizens can receive necessary medical care. Public medical insurance covers the majority of medical services that insurance systems provide to citizens, and the medical insurance claims dataset potentially enables population-scale prediction. In particular, this work attempts to exploit large language models (LLMs) to realize the future demand prediction from medical insurance claims. However, medical insurance claims are semi-structured and recorded with specific codes. The naïve use of LLMs offers poor prediction accuracy. To this end, we propose a multimodal representation for code-based MICs and a vision-language model for predicting medical demand. The experiments confirmed that our model achieved a significant improvement of 3.8 points in accuracy compared to the baseline system on diabetes cases.