Trustworthy AI for personalized glycemic control: a systematic review and critical appraisal of multimodal...

Tight glycemic control reduces acute complications (hypoglycemia, hyperglycemia) and long-term microvascular/macrovascular risks. The application of Artificial Intelligence (AI) to support optimal blood glucose management appears promising, however, its adoption at the bedside is still restricted. A recurring source of confusion is that short-horizon glucose forecasting and automated closed-loop control are often discussed together despite different evidence standards and safety requirements. In addition, trustworthy-AI practices (e.g., calibration/coverage, uncertainty-aware deferral, physiology-aware constraints, and reinforcement-learning safety with off-policy evaluation) are inconsistently reported. We addressed three questions: (RQ1) Which AI approaches provide reliable 30/60-minute glucose forecasts across datasets, modalities, model families, and validation rigor? (RQ2) Which control strategies improve time-in-range and safety versus established baselines, and at what evidence level? (RQ3) How frequently are trustworthy-AI elements reported or implemented, and where are the main gaps? We conducted a PRISMA-style systematic review (Jan 1, 2020-Jan 31, 2026) across MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, ACM Digital Library, and Cochrane. Dual screening and extraction were performed with inter-rater agreement quantified by Cohen's κ. We extracted setting, dataset/simulator, horizon/endpoints, modality, model/controller family, evaluation design, and trust/safety features; assessed risk of bias (PROBAST/QUADAS-2 where applicable); and summarized reporting alignment using a compact TRIPOD-AI snapshot. Synthesis was stratified by dataset, horizon, modality, architecture, and validation/evidence tier; quantitative synthesis was restricted to directly comparable strata; inverse-variance weighted descriptive summaries were used for strata with three to four studies, while formal random-effects meta-analysis was reserved for strata with at least five comparable effects. 80 studies met the inclusion criteria (forecasting: N = 50; control: N = 30). In Track A, 30/60-minute forecasts primarily served as short-term hypoglycemia alerting and near-term decision-support windows. A weighted descriptive summary was feasible only for a directly comparable OhioT1DM 30-minute RMSE stratum (k = 3), yielding RMSE 18.26 mg/dL (95% CI: 17.36-19.15), with heterogeneity not interpreted because of the small number of studies. In Track B, short prediction/control horizons functioned mainly as planning/action windows for controller decisions. The only quantitative Track B summary was a simulator-only UVA/Padova-style time-in-range (TIR) 70-180 mg/dL stratum (k = 3), yielding 76.42% (95% CI: 66.36-86.48; [Formula: see text] = 91.0%), but this was interpreted descriptively rather than as a common treatment effect because heterogeneity was substantial. Across both tracks, evidence was limited by heterogeneous validation protocols, sparse calibration/uncertainty reporting, and limited external validation; clinically validated model predictive control (MPC)/heuristic closed-loop comparators provided stronger translational evidence than predominantly simulator-only reinforcement learning (RL) controllers. Separating forecasting from control clarifies which gains are clinically established versus preclinical and supports a practical minimum checklist for reporting, evaluation rigor, and safety mechanisms to enable trustworthy translation.