AI framework to identify digital biomarkers of glycemic control and patient-reported outcomes for personalized diabetes management - SUMMARY With more than 500 million people affected worldwide and the number of cases steadily rising, diabetes presents a significant public health challenge. Thus, there is an urgent need for developing effective strategies for diabetes management to mitigate the burden of the disease on individuals and healthcare systems. Because most of the day-to-day care in diabetes is handled by patients, self-management is critical to maintain adequate glycemic control and prevent diabetes-related complications. However, self-management is a complex and involves continuously monitoring glucose levels, adhering to recommended treatment and medications, and adjusting diet and physical activity. The increasing availability of digital technologies and advancements in artificial intelligence (AI) offer new opportunities to improve decision support systems by enabling the monitoring and analysis of multi-modal health to identify digital biomarkers that provide insights into the various factors influencing diabetes, leading to a more comprehensive understanding of the disease and its management. Previous research works have focused on identifying digital biomarkers of glycemic control from glucose management data (i.e., glucose, insulin, nutritional, and physical activity data). The goal of this project is to expand the scope of past studies by exploring digital biomarkers that can capture clinically relevant glucose metrics, physiological states, and individuals’ behaviors to predict glucose and patient-reported outcomes. By leveraging physiological signals such as heart rate, step count, sleep data, glucose management data, and patient-reported outcomes from individuals with type 1 and type 2 diabetes, we aim to develop an AI-based framework to automatically identify biomarkers that can be used to better understand diabetes and its management and to inform the development of patient-centered decision support tools to both simplify diabetes management and improve glycemic control. This 2-year project involves collaboration between data scientists and clinical investigators to complete 2 proposed aims. Aim 1 involves (1.1) engineering biomarkers from multi-modal health data that relate to glycemic control, individuals’ behaviors, and patient-reported outcomes; (1.2) developing neural network-based data fusion models to predict glucose outcomes and patient-reported outcomes (e.g., sleep quality, fear of hypoglycemia, and diabetes distress surveys scores) using multi-modal digital biomarkers; and (1.3) using explainable AI and domain knowledge to identify the most relevant biomarkers. In Aim 2, (2.1) we will develop a random forest-based algorithm for building patient-centered decision support tools to provide actionable recommendations on what a person with diabetes can do to improve glycemic control considering the user’s specific goals and preferences. (2.2) We will demonstrate our approach in type 1 and type 2 diabetes in silico.
