AIDANET fully closed-loop algorithm

AIDANET is here because it targets the thing every commercial hybrid loop still struggles with: meals. The algorithm is being tested in fully closed-loop and hybrid modes, which makes it a useful research benchmark for how far insulin-only automation can go before insulin speed, sensing lag and missing meal context become unavoidable. It is not a product someone can choose today.

A July 2026 systematic review of meal-announcement-free algorithms — 69 studies, the class AIDANET belongs to — puts that promise in perspective: across the field, algorithms detected meals with a median sensitivity of 88% and precision of 93%, but only 25-40 minutes after eating began, and much of the evidence comes from in-silico simulation rather than people living their lives. That is the bar AIDANET's at-home trials have to clear with real data.

Registry freshness. NCT07039617 now lists active, not recruiting, with an estimated July 2027 completion after a September 2026 registry update; NCT06041971 still lists active, not recruiting with an old estimated completion. These are registry labels, not confirmation that a site is recruiting today or that results exist.

Accuracy is not safety. "Safe and Explainable Blood Glucose Forecasting" is the promise in the title of a September 2026 study, and its results make the distinction sharp: across two large retrospective datasets (848 participants), a physiology-constrained model called PhyNet — "A Physiology-Constrained Monotonic Neural Network for Safe and Explainable Blood Glucose Forecasting" — matched standard deep-learning models on 30-minute prediction: "predictive accuracy was similar across models (RMSE: 19.39-21.00 mg/dL; Time Gain: 10.45-13.35 min)." Yet "only PhyNet consistently captured the physiological effects of carbohydrates and insulin, yielding 0% unsafe recommendations versus up to 64.3% for baselines."⁷ The authors' conclusion is the point this gap keeps making: "Standard DL models can achieve state-of-the-art performance while failing to respect physiology, posing clinical risk. PhyNet preserves accuracy while enhancing physiological fidelity."⁷ The study also "generated counterfactual explanations to identify model-recommended actions for avoiding adverse events" — an early example of the explainability this gap asks for, alongside the safety envelope.⁷ Read it as evidence for the safety-envelope research direction above — a "30-minute horizon" forecasting result, not a trial of a controller dosing from these forecasts.⁷