Hypoglycemia remains one of the most feared acute complications for people with diabetes, especially those on intensive insulin therapy. A new deep learning model developed by an international consortium led by the University of Cambridge may soon change how we anticipate and prevent low blood glucose events. Published this week in The Lancet Digital Health, the study evaluated a recurrent neural network (RNN) trained on continuous glucose monitoring (CGM) data from 3,200 patients with type 1 diabetes across seven centers in Europe and Asia.
The model uses only three input streams: glucose values from the previous 60 minutes, meal timing (recorded via a mobile app), and insulin delivery data (from pump or injection logs). It does not require heart rate, accelerometry, or other wearable signals. The primary outcome was prediction of any hypoglycemic event (<70 mg/dL) occurring within the next 90 minutes. In the validation cohort of 800 patients, the model achieved 89% sensitivity (95% CI 86–92%) and 86% specificity (95% CI 83–89%), with a false alarm rate of 0.8 per patient per day.
More importantly, the model provided an average lead time of 54 minutes before the actual glucose value crossed the hypoglycemic threshold, giving patients enough time to consume fast-acting carbohydrates. In a simulated “alarm and treat” scenario, 78% of predicted hypoglycemia events were successfully aborted.
Dr. Elena Marchetti, lead author, said: “This is the first model that balances accuracy and simplicity. Because it works with CGM data alone, it can be deployed on a smartphone without cloud dependency.” SweetCare has already partnered with the Cambridge team to adapt the algorithm for its AI engine. A pilot version will be available to premium users in Q3 2026, with offline inference support for areas with poor internet connectivity.
The study also highlighted limitations: the model’s performance drops to 71% sensitivity in patients with highly irregular meal schedules (e.g., shift workers). Future work will incorporate real-time activity data from smartwatches to improve generalizability.