FDT-BioTech: Uncertainty Quantification in Deep Learning-Driven Digital Twins for Risk-Averse Decisions: Application in Type 1 Diabetes Management

Suny At Buffalo, Amherst NY

Investigators

Abstract

For individuals with Type 1 diabetes, keeping blood sugar levels within a safe range is essential but often difficult due to daily changes in diet, stress, activity, and other factors. This project aims to improve diabetes care by creating virtual models of individual patients--called digital twins--that can learn from wearable health sensors and help guide real-time insulin delivery using automated medical devices. By combining personal health data with artificial intelligence (AI), the project seeks to reduce the burden of self-management, prevent life-threatening highs and lows in blood glucose concentration, and improve long-term health outcomes. A central innovation of this work is the use of mathematical methods to quantify and manage uncertainty in predictions and recommendations made by AI models, thereby improving the reliability of treatment decisions. These methods also contribute to federal efforts to advance science for medical devices, supporting the safe and effective deployment of AI-empowered healthcare technologies. The broader impacts of the project include reducing diabetes-related complications and healthcare costs, improving public trust in AI-powered systems, and fostering interdisciplinary education. It will create educational opportunities for students and early-career researchers at the intersection of mathematical science, computational engineering, and biomedical science, including hands-on workshops in computational medicine, offering early exposure to high-performance computing and digital health technologies. This project develops new mathematical methods and computational algorithms to enable safe, reliable deployment of digital twins in healthcare, with a focus on managing Type 1 diabetes through personalized insulin delivery. It tackles three critical barriers that limit the trustworthiness of deep learning-based digital twins for healthcare, including identifying reliable models capable of accurately representing individualized glucose-insulin dynamics, quantifying predictive uncertainty under data scarcity, patient variability, and sensor errors, and validating treatment recommendations made by deep learning models under physiological fluctuations and potential control system faults. To close these gaps, the research advances three integrated technical objectives. First, it introduces an iterative Bayesian model selection and validation strategy for discovering deep learning models with accurate and reliable predictions, using population-level clinical data. Second, it implements algorithms and scalable cyberinfrastructure for real-time adaptation of the digital twin to individual physiology, including risk-averse insulin control. Third, it establishes rigorous methodologies for validating treatment recommendations by the deep learning-based digital twin, using both in silico simulations and clinical datasets. Intellectual contributions include a mathematical and computational framework for decision-making under uncertainty in physiological modeling, derivation of a posteriori error bounds for deep learning forecasts, and scalable techniques for optimal control under high-dimensional uncertainty. The resulting methods provide a generalizable blueprint for constructing, evaluating, and de-risking digital twins in a wide range of biomedical applications beyond diabetes care. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.