Artificial intelligence in glycemic management for diabetes: Applications, opportunities and challenges

Background

Diabetes mellitus, affecting 10% of the global population, stands as one of the most prevalent chronic diseases worldwide. Over half of these patients experience inadequate glycemic control, highlighting the shortcomings of conventional management methods. Given the economic and personal impact of diabetes, including decreased productivity, higher healthcare costs, severe complications and shortened lifespan, there is a pressing need for more efficient and cost-effective glycemic management tools. Nowadays, artificial intelligence (AI), which harnesses the power of big data and sophisticated algorithms, has offered promising solutions to improve patient outcomes and reduce the burden on healthcare providers. In this context, we demonstrate the current applications of AI in glycemic management and discuss the opportunities and challenges that AI faces in diabetes management.

Applications of AI in glucose monitoring and prediction

Glucose monitoring plays a pivotal role in diabetes management, encompassing hemoglobin A1C (A1C) measurement, blood glucose monitoring (BGM) by capillary (finger-stick) devices, and continuous (interstitial) glucose monitoring (CGM). AI has been employed in glucose monitoring and prediction to provide an early indicator of glycemic control status and detection of adverse events in patients with diabetes.

A1C, which reflects average glucose over the last 2–3 months, has been validated as a measure of both long-term glycemic control and risk for the development of complications. Several studies have constructed AI-driven HbA1c predictive models to forecast individual treatment outcomes and reflect the prognosis of diabetes, showing the potential of AI methods in long-term diabetes management.^[1]

Although A1C is significant, it does not adequately reflect glycemic variability and patterns, making it insufficient for patients requiring immediate monitoring and treatment adjustments. In contrast, BGM and CGM provide more detailed and timely glucose response data. Capillary blood glucose monitoring is widely used to assess glucose levels and guide both self-and clinician-directed management. CGM allows continuous, real-time monitoring of interstitial glucose, offering insights into glycemic excursions, daily profiles, and variability. AI algorithms based on BGM could predict future BG and adverse events, providing early warnings of hyperglycemia or hypoglycemia.^[2] Notably, ongoing improvements in the accuracy and convenience of CGM devices have prompted the increasing adoption of this technology. Several studies have utilized the combined data from CGM and other wearable devices like physical activity monitors and meal tracking to understand the precise glucose change of patients deeply and further support personalized treatment.^[3]

AI used in optimizing diabetes management

While traditional diabetes management confronts challenges such as poor patient adherence, lack of personalized care, and limited medical resources, there are increasing applications of AI in diabetes education, lifestyle guidance, and pharmacologic therapy, offering a promising solution to provide optimal diabetes treatments.

Opportunities and challenges

The advancement of AI technology enables more precise monitoring and prediction of short-term and long-term blood glucose levels and offers personalized lifestyle guidance, anti-diabetic medication, and insulin decision-making, providing new tools to improve glycemic control in diabetic patients and reduce the risk of complications (Figure 1).

Figure 1

Applications of artificial intelligence (AI) in glycemic management for diabetes.

However, AI applications in diabetes management still face challenges. First, low-quality or small datasets can result in inadequate model performance and biases, harming clinical practice. Collaborative efforts to establish standardized data collection platforms are crucial for overcoming this obstacle. Furthermore, the emergence of wearable devices and the Internet of Things (IoT) promises to yield a wider range of high-precision data. However, as the gathered data may encompass sensitive personal details, data privacy remains a paramount concern in healthcare AI. To safeguard privacy, regulations like the General Data Protection Regulation (GDPR) and California Consumer Privacy Act (CCPA) have been enacted. Simultaneously, innovative techniques such as federated learning (FL), blockchain technology, and generative adversarial networks are evolving to tackle this challenge. Second, while high-performance deep learning models are powerful tools, they often function as “black boxes”. Clinical practice requires interpretability from these models to ensure patient safety and foster trust among physicians. Hence, there is an urgent need for algorithmic innovation to construct high-performance transparent models or advanced interpretability algorithms to further expand the applicability of models. Third, a notable disparity exists between laboratory environments and actual clinical situations, and current evaluations of models are still inadequate. Consequently, during the implementation of AI, it is necessary to follow guidelines for step-by-step validation and iterate and optimize models based on real-world requirements to ensure their safety and effectiveness. Simultaneously, ethical questions arise regarding accountability in the event of AI errors, highlighting the importance of robust safety measures and unambiguous policies assigning responsibility in AI applications (Figure 2).

Figure 2

Challenges for artificial intelligence (AI) application in glycemic management and possible strategies.

In conclusion, the advancement of AI is transforming diabetes management yet several challenges remain to be addressed. As physicians, we should not only embrace and refine these technologies to offer better healthcare services but also keep a cautious, patient-centered orientation to ensure safety.