Summary: In 149 hospitalized adults with type 2 diabetes (HbA1c 7.0 to 11.0 percent) across three Chinese centres, an artificial intelligence insulin titration decision-support system (iNCDSS) was noninferior to senior endocrinology physicians for proportion of time in the target glucose range of 70 to 180 mg/dL over five days (76.4 percent versus 73.6 percent; difference 2.7 percentage points, 95 percent CI -2.7 to 8.0; noninferiority margin 6 points). Adverse events did not differ significantly between groups. This establishes feasibility and noninferiority over a short inpatient window, not superiority or durable benefit.
PICO Summary
| Element | Detail |
|---|---|
| Population | 149 hospitalised adults (mean age 64.2 years, 56.4 percent male) with type 2 diabetes and HbA1c 7.0 to 11.0 percent, on antidiabetic treatment in the prior 3 months; multicentre, single-blind, parallel randomised controlled trial across 3 endocrinology wards in China (2021 to 2022). |
| Intervention | Real-time AI-based insulin clinical decision-support system (iNCDSS) directing insulin dosage titration for 5 consecutive days (n=75). |
| Comparison | Insulin dosage titration by senior endocrinology physicians (standard care) for 5 consecutive days (n=74). |
| Outcome | Primary outcome, proportion of time in target glucose range (70 to 180 mg/dL): iNCDSS 76.4 percent (SD 16.4) versus physician 73.6 percent (SD 16.8); estimated treatment difference 2.7 percentage points (95 percent CI -2.7 to 8.0), meeting the prespecified noninferiority margin of 6 percentage points. No significant between-group difference in adverse events. No ARR or NNT applicable (continuous noninferiority endpoint). |
AI insulin titration vs physicians
RCT · type 2 diabetes · 5 days
AI insulin titration was noninferior to senior physicians for time in range over five inpatient days; the difference was small and its CI sat within the prespecified margin. This shows feasibility, not superiority or durable benefit.
Expert Commentary
This is a clean, prespecified noninferiority randomised trial, and on its own terms it succeeds: the AI system matched senior endocrinologists for time in range over five inpatient days, and the confidence interval sat comfortably within the six-point margin. The verdict should be read narrowly. Noninferiority is not superiority, and the upper bound of the interval still permitted a modest physician advantage, so the honest claim is that the system did not perform meaningfully worse, not that it dosed better. The central limitation is the setting and horizon. Five days on a monitored endocrinology ward, with HbA1c capped at 11 percent and a relatively small sample, tells us little about ward-to-discharge transitions, hypoglycaemia over weeks, or performance where nursing and laboratory cadence is less tightly controlled. The trial was single-blind, meaning treating clinicians knew the allocation, which can subtly shape co-interventions and threshold decisions; the physician satisfaction data, while encouraging, are inherently unblinded and should not be read as an efficacy signal. Can I use this with my patients? Not yet. The most defensible near-term use is as a supervised decision aid for inpatient titration in well-resourced endocrinology units, pending external validation, longer follow-up, and hard hypoglycaemia outcomes. I would like to see a multinational trial with blinded outcome adjudication and a clear safety endpoint before this tool moves from promising to practice-changing.
References
Ying Z, Fan Y, Chen C, et al. Real-Time AI-Assisted Insulin Titration System for Glucose Control in Patients With Type 2 Diabetes: A Randomized Clinical Trial. JAMA Netw Open. 2025;8(5):e258910. doi:10.1001/jamanetworkopen.2025.8910