Artificial intelligence (AI) is rapidly reshaping diabetes care, expanding its role in complication screening, risk stratification, insulin optimization, and hospital operations, according to a review published in the journal Endocrine Practice.
With 537 million people worldwide living with diabetes and a projected number of 783 million by 2045, the burden of the disease continues to rise. Researchers investigated how AI and machine learning (ML) are being integrated into clinical workflows to improve early detection, personalize treatment, and streamline healthcare delivery.
“Artificial intelligence has the potential to significantly enhance diabetes management by enabling proactive risk prediction, optimizing insulin dosing, personalizing treatment plans, and improving clinical decision-making across the treatment continuum,” said Rohit Parab, M.D., of the Department of Endocrinology at Emory University School of Medicine in Atlanta, and colleagues.
Screening and early detection
AI-powered imaging tools are advancing screening for diabetic retinopathy (DR), neuropathy, and foot ulcers. A machine learning model analyzing retinal images demonstrated 93% sensitivity and 91% specificity for detecting DR. Three fully automated systems have received U.S. Food and Drug Administration clearance for DR diagnostics: IDx-DR (now LumineticsCore), EyeArt, and AEYE Diagnostic Screening.
Beyond retinopathy, ML approaches have also been applied to detect diabetic peripheral neuropathy. Algorithms that analyze thermograms, skin images, and corneal confocal microscopy can identify early neurological changes. The review authors noted that some models use only clinical features to estimate the severity of neurological deficits, which could reduce reliance on specialized testing.
Digital biomarkers represent another emerging field. One study predicted interstitial blood glucose levels with up to 87% accuracy by combining smartwatch data and food records. A wristband-based photoplethysmography model detected diabetes and prediabetes with more than 84% accuracy. Smartphone-based vascular signal analysis also demonstrated feasibility for diabetes screening.
Risk stratification and disease progression
AI-based risk engines are being developed to predict progression from prediabetes to diabetes, cardiovascular disease outcomes, and other complications. Prediction performance varied depending on the model and input variables, with reported area under the curve values ranging from 0.64 to 0.93.
The BRAVO (Building, Relation, Assessing, and Validating Outcomes) diabetes model integrates 17 interrelated risk equations to simulate disease progression and complications. The model has been calibrated against randomized controlled trial data and national survey data and is implemented within the electronic health record.
Machine learning models integrating genetic data, polygenic risk scores, imaging, and electronic medical record data have further enhanced genome-based risk prediction, with reported area under the curve values ranging from 0.61 to 0.94.
Automated insulin dosing and blood sugar optimization
Wearable technologies such as continuous glucose monitoring (CGM) and automated insulin dosing (AID) systems are central to AI-enhanced management. Although commercially available AID systems primarily rely on deterministic control algorithms, AI-based enhancements are also emerging.
In a multicenter pilot study in pediatric patients aged 2 to 6 years with type 1 diabetes, an AI-driven digital twin approach optimized pump settings and reduced delivery times over 8 weeks. The neural net artificial pancreas system achieved up to 85% of the time in range in hybrid closed-loop mode and 75% of the time in fully closed-loop mode during a supervised short-term study and was subsequently validated in a home setting.
Decision support systems also assist with insulin dose adjustment. A voice-based conversational AI application reduced time to optimal basal insulin dosing, adherence, glycemic control, and diabetes-related psychological distress in patients with type 2 diabetes compared to standard care. In young people with type 1 diabetes, an AI-based insulin dose optimization system demonstrated non-inferiority to physician-directed dose adjustment.
Researchers note that an ML model applied to inpatient electronic medical record data predicted insulin requirements more accurately than traditional guidelines.
Patient self-management and dietary analysis
AI-powered smartphone interventions have been demonstrated to improve blood glucose outcomes. In one trial, a smartphone-based behavioral coaching intervention reduced hemoglobin A1c by 1.9% compared to 0.7% in the control group. The AI-powered dietary management platform was also associated with improved hemoglobin A1c and greater weight loss compared to routine care.
Deep learning systems that analyze images of food can identify food types, portion sizes, and estimate calories. Dr. Parab and colleagues reported a new model that integrates dietary image analysis with historical blood sugar data and contextual factors such as physical activity and sleep to predict postprandial blood sugar fluctuations and alert patients to potential hyperglycemic or hypoglycemic events.
Administrative applications and implementation challenges
AI integration extends beyond direct patient care. Applications include resource allocation, staff scheduling, billing automation, electronic medical record documentation, and more. In a cross-sectional study, discharge summaries generated by a large-scale language model were comparable in overall quality to physician-generated summaries, but errors were present and potential harm was rated lower.
Despite these advances, challenges remain. Model performance varies based on input variables and datasets, and limited demographic representation can introduce bias. Additional concerns include privacy, data security, interoperability, regulatory standards, and the need for standardized assessment frameworks.
Efforts to advance responsible AI include explainable models that increase transparency, frameworks for unbiased evaluation of retinal screening systems, and privacy-preserving federated learning approaches that enable multicenter training without direct data sharing.
The researchers reported multiple disclosures that can be found in published studies.
