The global burden of diabetes is growing, with more than 800 million people currently affected and projections indicating over 1.3 billion cases by 2050. More than 90% are attributed to type 2 diabetes (T2D), underscoring the need for approaches that support timely, personalised care. Artificial intelligence and wearable technology are increasingly aligned to address this need. Continuous glucose monitors (CGMs), smartwatches and other sensors collect real-time physiological data, while AI models interpret these streams to forecast glycaemic trends, inform insulin adjustments and guide self-management. A recent review examined how AI and wearables intersect in clinical and self-management contexts, identifying advances in glycaemic monitoring, adaptive insulin support and event prediction, alongside persistent challenges in demographic representation, data quality, benchmarking and model interpretability.
Expanding Data from Continuous Monitoring
Wearables have reshaped glucose monitoring by providing continuous, minute-by-minute insight into interstitial glucose and related physiology. In the surveyed studies, CGMs featured in 70% of projects, reflecting their central role in capturing real-time glycaemic dynamics that support day-to-day decisions on insulin, diet and physical activity. A further 20% used fitness trackers or smartwatches to quantify activity, heart rate and other metrics relevant to metabolic control. Less common but notable were photoplethysmography and electrodermal activity devices, which together accounted for about 10% of investigations. These multimodal data streams extended beyond glucose levels to include heart rate variability, sleep patterns and other parameters that can contextualise glycaemic fluctuation.
Must Read: Tech-Driven Shift in Diabetes Monitoring
The scope of outcomes assessed has broadened over the past decade. Early work focused largely on short-horizon glucose prediction, while later studies diversified into insulin management and classification tasks. Additional objectives included detecting physical activity, estimating stress from wearable signals, evaluating diabetic retinopathy using CGM-derived features and examining the effect of CGM sensor placement on forecasting error. This diversification reflects growing confidence in continuous monitoring as a foundation for AI-enabled decision support that can adapt to everyday conditions.
Evolving AI Techniques and Clinical Uses
AI architectures in the review ranged from traditional machine learning to advanced deep learning tailored for time-series data. Recurrent neural networks and long short-term memory models were used in 45% of studies, capitalising on their ability to model temporal dependencies in wearable streams. Random forests, support vector machines and other established algorithms remained prevalent, offering baselines and, in some cases, simpler paths to interpretation. These models were applied to forecasting interstitial glucose at multiple horizons, classifying glycaemic states and detecting adverse events, with some studies integrating dietary or activity information to refine predictions.
Applications extended into adaptive therapy. Several projects paired AI models with insulin titration frameworks or broader decision support, linking CGM data and clinical parameters to dose recommendations or alerts. Others explored population-specific prediction, domain generalisation across datasets and the fusion of signals from multiple sensors to improve robustness. Across designs, inputs commonly included timestamped CGM readings enriched with behavioural or physiological features. Outputs ranged from near-term glucose forecasts to risk stratification for hypoglycaemia and hyperglycaemia, as well as indicators relevant to complications such as diabetic retinopathy. Although methodologies differed, the overarching pattern pointed to incremental gains in prediction accuracy and the potential to embed AI within routines of care and self-management.
Study designs were varied. Two thirds were observational or experimental, including 20% that prospectively captured real-world wearable data in naturalistic settings. A smaller proportion used randomised or non-randomised experimental designs to evaluate interventions involving CGMs and activity trackers. Sample sizes ranged from five to more than one thousand participants, with a median of 150, and 40% enrolled fewer than 100, which may constrain generalisability. Participants were predominantly adults with T2D, with an average age of 55 years and balanced gender distribution at around 48% female. Data collection settings spanned outpatient clinics, free-living environments and controlled studies, while some investigations used simulated cohorts to test algorithmic strategies under standardised conditions.
Gaps in Evidence and Priorities for Implementation
Despite progress, several limitations could hinder equitable and reliable deployment. Reporting on participant diversity was sparse, with only 7% of studies detailing racial or ethnic composition and limited representation of minority populations. Geographical distribution was concentrated, with 45% conducted in North America, 30% in Asia, 20% in Europe and 5% elsewhere, leaving regions such as Africa and South America under-represented. These imbalances raise questions about external validity when models are applied to health systems and populations that differ from development cohorts.
Data quality and benchmarking also emerged as recurring issues. Differences in sensor accuracy, sampling intervals and noise handling complicate cross-study comparisons. The absence of standardised evaluation frameworks makes it difficult to assess whether performance gains reflect methodological innovations or dataset idiosyncrasies. While deep learning models delivered practical improvements, limited interpretability reduced transparency for clinicians and patients. Few studies explicitly addressed algorithmic bias or provided mechanisms to explain model outputs in ways that support shared decision-making. Future work identified by the review prioritised clearer performance benchmarks, improved model transparency and targeted efforts to address demographic disparities through broader recruitment and reporting.
AI integrated with wearables is advancing T2D management by enabling continuous monitoring, more accurate forecasting and support for insulin adjustment and self-care. Evidence from 60 studies highlights momentum across prediction, classification and decision support, with CGMs as the dominant data source and complementary signals from activity and physiological sensors augmenting context. Real-world designs are increasingly common, yet gaps in demographic diversity, data standards and interpretability remain. Addressing these issues through inclusive study populations, transparent modelling and agreed evaluation benchmarks will help clinicians, researchers and technology developers translate promise into equitable outcomes for people living with diabetes.
Source: npj digital medicine
Image Credit: iStock
