Predicting hospitalization with LLMs from health insurance data.

Area of Science:

• Health Informatics
• Machine Learning
• Natural Language Processing

Background:

• Health insurance data holds untapped potential for predicting hospitalizations.
• Extracting features from this data for machine learning traditionally requires specialized expertise.
• Large language models (LLMs) offer a new avenue to leverage this data with reduced technical barriers.

Purpose of the Study:

• To present an approach for utilizing health insurance data with LLMs to predict hospitalizations.
• To develop and evaluate pre-trained LLMs for hospitalization prediction in Portuguese and English.

Main Methods:

• Organizing and preparing health insurance data for LLM input.
• Pre-training LLMs on curated health insurance datasets.
• Evaluating model performance on predicting general hospitalizations and stroke-related hospitalizations.

Main Results:

• Generated pre-trained LLMs in Portuguese and English capable of predicting hospitalizations.
• Achieved high performance metrics: F1-Score = 87.8 and AUC = 0.955 for general hospitalization prediction.
• Attained excellent results for stroke-related hospitalization prediction: F1-Score = 88.7 and AUC = 0.964.

Conclusions:

• LLMs can effectively predict hospitalizations using health insurance data, democratizing access to predictive health applications.
• The developed pre-trained models demonstrate significant accuracy and potential for diverse healthcare applications.
• Models are publicly available to the scientific community to foster further research and development.