Author: Bingyu Mao, MA (2025)
Primary advisor: Degui Zhi, PhD
Committee members: Ziqian Xie, PhD, Masayuki Nigo, MD and Hua Xu, PhD
PhD thesis: McWilliams School of Biomedical Informatics at UTHealth Houston.
ABSTRACT
Advancements in precision medicine increasingly rely on data-driven approaches to improve clinical decision-making. This work presents three interconnected studies that leverage deep learning, reinforcement learning, and comparative modeling to address key challenges in precision drug dosing and disease risk prediction. Together, these contributions offer novel computational frameworks that enhance model performance, improve dosing strategies, and guide model selection for clinical predictive tasks.
To improve individualized vancomycin therapeutic drug monitoring (TDM), the first study introduces PKRNN-2CM, a novel deep learning framework that integrates a two-compartment pharmacokinetic (PK) model with recurrent neural networks (RNNs). While one-compartment models are commonly used in clinical practice, the two-compartment model better reflects vancomycin PK. PKRNN-2CM was trained on both simulated and real-world electronic health record (EHR) data, demonstrating improved accuracy over the simpler PKRNN-1CM model in predicting vancomycin concentration trajectories. In real-world data evaluations, PKRNN-2CM showed statistically significant improvements in prediction accuracy and more precise estimates of the area under the concentration-time curve (AUC), a clinically important metric for dose optimization. These findings highlight PKRNN-2CM's potential to support individualized vancomycin dosing strategies and improve patient outcomes.
The second study introduces a reinforcement learning-based simulation framework designed to evaluate vancomycin dosing strategies under realistic clinical conditions. This framework employs PKRNN-2CM to generate vancomycin concentration-time curves and incorporates a novel AUC reward score to assess the effectiveness of various dosing prescriptions. Results indicate that targeting an AUC of 450 mg·h/L leads to a higher proportion of dosing prescriptions achieving the therapeutic range, especially when accounting for noise. The framework’s ability to simulate clinical scenarios provides a valuable tool for improving the early achievement of therapeutic vancomycin levels and paves the way for refining deep learning models for dynamic dosing recommendations.
The third study investigates disease risk prediction by benchmarking the performance of specialized clinical foundation models (CFMs), such as Med-BERT, against general-purpose large language models (LLMs). Classifiers built on these models were evaluated for predicting the risk of pancreatic cancer and heart failure among diabetic patients. Results indicated that CFMs significantly outperformed LLMs in tasks involving large and complex EHR and claims datasets. However, on datasets with limited patient information, fine-tuned LLMs achieved slightly better performance, with no statistically significant difference. These findings emphasize the importance of aligning model selection with data characteristics and clinical objectives when deploying AI-driven disease risk prediction models.
Collectively, this work advances precision medicine by introducing novel deep-learning frameworks, innovative evaluation methodologies, and practical insights for model selection. The PKRNN-2CM model enhances predictive accuracy for vancomycin precision dosing, while the reinforcement learning-based framework provides a data-driven strategy for optimizing clinical guidelines for dosing recommendations. Additionally, the comparative evaluation of CFMs and LLMs informs future model deployment in clinical predictive tasks. These contributions provide methodological advancements for biomedical informatics research and offer practical tools to improve clinical decision-making and patient outcomes.