Medical-AI trial: AI-driven insulin dose adjustment vs. clinicians — glycemic control in type 2 diabetes on general wards (multicenter RCT) (NCT06319300)

For inpatients with diabetes, insulin doses must be adjusted frequently to match blood glucose. Too little leaves glucose high; too much causes hypoglycemia, a dangerous state. This adjustment leans heavily on clinician experience and effort, and wards face constraints of staffing and time. This study examines whether an AI that proposes insulin doses from the glucose trajectory can be used here.

Per the registry summary, this is a single-blind, multicenter randomized controlled trial (RCT) in type 2 diabetes patients on general wards requiring subcutaneous insulin. Participants were randomized 1:1: one group adjusted doses with an AI-assisted insulin decision system, the other by clinician instruction as usual.

The design then compares glycemic control and adverse-event risk between the groups to assess the effectiveness and safety of the system. The primary outcome is the proportion of time blood glucose stayed in target range (3.9–10.0 mmol/L), known as time in range (TIR). The planned enrollment is about 142, and the study is completed.

Insulin adjustment is a delicate task threading a narrow safety margin between too much effect (hypoglycemia) and too little (hyperglycemia), and it is hard to standardize. If AI can propose doses objectively, it might reduce operator-to-operator variation and ease ward workload.

Comparing head-to-head against clinician adjustment in an RCT and evaluating not only effectiveness but also safety such as hypoglycemia is a meaningful validity check for embedding medical AI into an actual treatment process. That said, this article presents the design and primary outcome; it does not state which was superior or opine on clinical adoption.