Predicting Patient Health Pathways Using Large Language Models for Digital Twins
Overview
The Digital Twin—Generative Pretrained Transformer (DT-GPT) leverages large language models to predict longitudinal clinical trajectories from electronic health records without data imputation. Benchmarking on datasets for non-small cell lung cancer, ICU patients, and Alzheimer’s disease demonstrated DT-GPT’s superior forecasting accuracy and explainability compared to state-of-the-art models.
Background
Clinical forecasting is critical for monitoring patient outcomes, guiding treatment decisions, and supporting drug development. Digital twins—virtual patient representations—enable personalized simulation of health trajectories by integrating diverse longitudinal data. Generative AI, particularly large language models (LLMs), offers promising capabilities for creating digital twins but faces challenges such as handling missing data and modeling complex variable interactions. DT-GPT addresses these challenges by fine-tuning LLMs on real-world electronic health record data to forecast multivariate clinical trajectories with improved accuracy and interpretability.
Data Highlights
Dataset
Scaled Mean Absolute Error Reduction
Non-Small Cell Lung Cancer
3.4%
Intensive Care Unit
1.3%
Alzheimer’s Disease
1.8%
Key Findings
DT-GPT forecasts clinical variable trajectories across short-, medium-, and long-term horizons using EHR data without requiring imputation or normalization.
It outperforms state-of-the-art machine learning models by reducing scaled mean absolute error by up to 3.4% across diverse clinical datasets.
DT-GPT maintains the distributions and cross-correlations of clinical variables, capturing complex interdependencies inherent in patient data.
The model demonstrates zero-shot forecasting capabilities, enabling prediction of clinical variables without prior task-specific training.
Explainability is enhanced through a human-interpretable chatbot interface, facilitating clinical insight into model predictions.
Clinical Implications
DT-GPT’s ability to accurately predict longitudinal patient trajectories supports personalized medicine by enabling real-time monitoring and treatment optimization. Its zero-shot forecasting and interpretability features may facilitate clinical decision-making and adverse event mitigation without extensive retraining. This approach paves the way for integrating digital twins into clinical trials and routine care to improve patient outcomes.
Conclusion
DT-GPT represents a significant advancement in clinical forecasting by harnessing large language models to create interpretable digital twins that accurately predict patient health pathways. Its robust performance across multiple diseases highlights its potential as a versatile tool for precision medicine applications.
References
Author/Source/2024 -- Predicting Patient Health Pathways through Large Language Models for Digital Twin Applications
