Leveraging Large Language Models to Predict Patient Health Trajectories for Enhanced Digital Twin Applications

DT-GPT is a novel framework that utilizes fine-tuned pre-trained large language models (LLMs) on clinical data to forecast patients’ laboratory values. This method is adaptable across various text-focused LLM architectures, making it versatile in its application. DT-GPT was tested on three datasets: non-small cell lung cancer (NSCLC), intensive care unit (ICU), and Alzheimer’s disease (AD). The framework transforms electronic health records (EHRs) into a compatible text format and employs BioMistral, a biomedical LLM, for enhanced predictive accuracy. Key laboratory values, including hemoglobin and leukocytes, were predicted based on patient histories to evaluate treatment responses. The study employs standardized metrics like mean absolute error (MAE) and area under the receiver operating characteristic curve (AUC) for performance assessment. Furthermore, DT-GPT effectively handles missing data and noise through rigorous preprocessing, establishing itself as a robust tool in clinical predictions and patient outcome forecasts.

Source link