AI-CARE: A Novel AI-Driven Mobile Insulin Calculator Integrating Image- Based Carbohydrate Estimation and CGM

Diabetes mellitus (DM) is a chronic metabolic disorder affecting an estimated 537 million adults worldwide, with prevalence projected to increase to 643 million by 2030. Poor glycaemic control is strongly associated with an increased risk of microvascular and macrovascular complications. Landmark trials, including the Diabetes Control and Complications Trial and the UK Prospective Diabetes Study, demonstrated that intensive glycaemic control achieved through multiple daily insulin injections (MDI) significantly reduces long-term complications. MDI typically consists of basal and bolus insulin therapy. Basal insulin is a long-acting insulin and is used to maintain background insulin levels and control glucose during fasting periods, including overnight and between meals. Bolus insulin, which is rapid-acting, is administered before meals to manage postprandial glucose excursions resulting from carbohydrate intake.

Clinical studies report that initiating MDI can reduce glycated haemoglobin (HbA1c) by approximately 1.5-2% within 3-4 months, with sustained reductions of up to 3-4% over 1 year. Epidemiological data indicate that each 1% reduction in HbA1c is associated with a 37% decrease in microvascular complications. Despite these benefits, a meta-analysis reported that only 23.75% of individuals with insulin-treated type 2 diabetes had HbA1c levels below 7%, while a cross-sectional study found that only 20% of individuals with type 1 diabetes attained the same target. Achieving optimal glycaemic control in individuals treated with MDI requires accurate pre- meal glucose assessment, precise carbohydrate estimation, and appropriate bolus insulin dose calculation. However, each of these components presents substantial real-world challenges.

Self-monitoring of blood glucose (SMBG) plays an important role in improving glycaemic control and empowering individuals with diabetes. It is particularly beneficial for those treated with insulin, as it helps determine bolus insulin dosing, facilitates insulin dose titration, and enables timely detection of hypoglycaemia. Despite strong recommendations for SMBG use among insulin-treated individuals, its uptake remains suboptimal in many Asian countries. In many Asian settings, blood glucose monitors and test strips are paid out of pocket, creating substantial financial barriers to regular testing. Furthermore, finger-prick testing is often perceived as painful and inconvenient. Patients may avoid monitoring due to frustration with high glucose readings, embarrassment when testing in public, fear of needles, workplace constraints, or limited motivation and self-efficacy. Collectively, these barriers result in inadequate pre-meal glucose monitoring, leading to inappropriate bolus dosing and suboptimal insulin titration, thereby contributing to poor glycaemic control despite insulin therapy.

Beyond glucose monitoring, carbohydrate counting-although a foundational component of MDI therapy-is cognitively demanding, time-consuming, and error-prone, particularly in multicultural dietary environments. In Singapore and other Asian settings, meals are commonly consumed in hawker centres, food courts, and home-based contexts where dishes are mixed, visually complex, and rarely accompanied by standardized nutritional information. Staple foods such as rice, noodles, congee, roti prata, and mixed-economy rice involve variable portion sizes and composite ingredients, making accurate carbohydrate estimation particularly challenging. Inaccurate estimation can lead to insulin underdosing or overdosing, resulting in postprandial hyperglycaemia or hypoglycaemia. Notably, mean carbohydrate-counting accuracy among individuals using MDI has been reported to be only 59%.

Even when glucose levels are measured and carbohydrate content estimated, bolus insulin calculation requires application of insulin-to-carbohydrate ratios (ICR), insulin sensitivity factors (ISF), and correction algorithms. This process demands numeracy and health literacy skills that are frequently limited among individuals with diabetes and are associated with poorer glycaemic outcomes. Difficulties in performing dose calculations may lead to reliance on fixed bolus doses, empirical estimation, incorrect administration, or omission of boluses. Indeed, 64% of patients have been shown to inaccurately estimate their prandial insulin requirements, and fixed-dose regimens have been associated with increased glycaemic variability and weight gain.

Collectively, these interrelated challenges-suboptimal glucose monitoring, inaccurate carbohydrate estimation, and numeracy burden in insulin dose calculation-contribute to postprandial glucose excursions and persistent hyperglycaemia despite insulin therapy. Addressing these barriers through integrated, automated clinical decision-support systems may therefore represent a promising strategy to enhance dosing accuracy, reduce cognitive burden, and improve overall glycaemic outcomes. AI-CARE is an artificial intelligence-enabled clinical decision support mobile application designed to support prandial insulin dosing in individuals receiving MDI, incorporating a "human-in-the-loop" approach to ensure user verification of recommended doses. The application integrates AI-driven image-based carbohydrate estimation tailored to Asian dietary contexts, individualized patient-specific insulin parameters, and real-time CGM data to generate personalized insulin dose recommendations.

The system comprises several key components. First, the image-based carbohydrate estimation module enables users to capture meal images at standardized angles (45-55° and 90°), after which the AI engine automatically estimates carbohydrate content. User-in- the-loop features, including portion adjustment and manual editing, are incorporated to enhance accuracy and allow verification. Second, real-time CGM data are integrated into the system, allowing current glucose values to be incorporated into the insulin dose calculation algorithm. Third, the application calculates personalized insulin dose recommendations based on parameters such as the ICR, ISF, and target glucose level, with predefined maximum bolus limits to enhance safety.

To ensure safe use, the application incorporates multiple safety features and feedback mechanisms, including dose caps, prompts to review glucose levels prior to insulin administration, and reminders to confirm dosing decisions. Although AI-CARE provides insulin dose recommendations, the final decision remains patient-driven, and users retain full autonomy to accept, modify, or override the suggested dose.

The primary objective of this study is to evaluate the feasibility, safety, and preliminary effectiveness of the AI-CARE mobile application among adults with diabetes receiving MDI. Feasibility will be assessed in terms of recruitment, retention, participant adherence, and application usability. Safety will be evaluated by monitoring the incidence of hypoglycaemia, hyperglycaemia, and other adverse events throughout the intervention period.

The study will also examine the preliminary effectiveness of AI-CARE on glycaemic and patient-reported outcomes compared with usual care. In addition, a qualitative process evaluation will be conducted to explore participants' experiences and perceptions of using AI-CARE, including perceived ease of use, usefulness, acceptability, and barriers and facilitators to real-world implementation. Findings from this phase will inform further refinement and optimization of the AI-CARE system and guide the design of a future large- scale randomized controlled trial.

This study adopts a mixed-methods design, comprising a 12-week feasibility pilot randomized controlled trial with an embedded qualitative process evaluation using focus group discussions.