Using Nationwide Health Records to Address Care Gaps in Diabetes and Prediabetes Diagnosis through AI Assisted Large-Scale Prioritized Outreach for Hemoglobin A1c Screening

Abstract Undiagnosed diabetes and prediabetes affect almost one in three adults in the United States. Early detection of dysglycemia can prevent disease progression and associated morbidity and mortality. Current guidelines recommend broad screening but they are not consistently followed. Artificial intelligence (AI) methods can be effective for diabetes prediction, but it remains unclear whether even a highly accurate predictive model guarantees improved health outcomes. In this study of over 7 million at-risk patients in the US between 2014 and 2020 we evaluated the potential for AI to improve diabetes screening effectiveness on the basis of two metrics: the fraction of tests recommended to patients that will ultimately test normal (wastefulness), and the fraction of patients with dysglycemia not detected from an eligible population (miss rate). Our results revealed that deep learning can identify over twice the number of patients as current clinical practice using no additional screening resources. This method could be developed into a large-scale, cost-effective prioritized screening program, allowing for early-stage interventions to prevent diabetes progression, and associated adverse health outcomes.