Large language models forecast patient health trajectories enabling digital twins

Area of Science:

• Artificial Intelligence
• Computational Biology
• Medical Informatics

Background:

• Generative AI is transforming digital twin technology, creating virtual patient models for health prediction.
• Large language models (LLMs) show significant potential for clinical forecasting applications.

Purpose of the Study:

• To develop and evaluate the Digital Twin-Generative Pretrained Transformer (DT-GPT) for clinical trajectory prediction.
• To extend LLM-based forecasting to complex healthcare data challenges.

Main Methods:

• Developed DT-GPT, an LLM extension for clinical forecasting.
• Utilized electronic health records (EHRs) without imputation or normalization.
• Benchmarked against state-of-the-art machine learning models on diverse datasets.

Main Results:

• DT-GPT outperformed existing models on non-small cell lung cancer, ICU, and Alzheimer's datasets, reducing scaled mean absolute error by 3.4%, 1.3%, and 1.8%, respectively.
• The model preserved clinical variable distributions and cross-correlations.
• Demonstrated explainability via a human-interpretable interface and zero-shot forecasting capabilities.

Conclusions:

• DT-GPT offers a robust solution for clinical trajectory prediction using EHRs, overcoming data limitations.
• LLMs show promise as platforms for clinical forecasting, with potential applications in trials, treatment selection, and adverse event mitigation.