Real-Time AI-Assisted Insulin Titration System for Glucose Control in Patients With Type 2 Diabetes: A Randomized Clinical Trial

Abstract

Importance: Type 2 diabetes (T2D) is one of the most prevalent chronic diseases in the world. Insulin titration for glycemic control in T2D is crucial but limited by the lack of personalized and real-time tools.

Objective: To examine whether an artificial intelligence-based insulin clinical decision support system (iNCDSS) for glycemic control in hospitalized patients with T2D is noninferior to standard insulin therapy administered by senior physicians.

Design, setting, and participants: This multicenter, single-blind, parallel randomized clinical trial (RCT) was conducted between October 1, 2021, and September 8, 2022, in endocrinology wards of 3 medical centers. Eligible participants were adults (aged ≥18 years) with glycated hemoglobin levels between 7.0% and 11.0% who had received antidiabetic treatments in the previous 3 months.

Interventions: Participants were randomized in a 1:1 ratio to receive insulin dosage titration by iNCDSS or senior endocrinology physicians for 5 consecutive days.

Main outcomes and measures: The primary outcome was the proportion of time in the target glucose range (70-180 mg/dL) during the 5-day study period; the noninferiority margin was 6 percentage points. Secondary outcomes included other glycemic control measurements and adverse events.

Results: A total of 149 participants (mean [SD] age, 64.2 [12.0] years; 84 male [56.4%]) were enrolled and randomized to the iNCDSS group (n = 75) or physician group (n = 74). The mean (SD) target glucose range (primary outcome) was 76.4% (16.4%) in the iNCDSS group and 73.6% (16.8%) in the physician group, which achieved the prespecified noninferiority criterion (estimated treatment difference, 2.7%; 95% CI, -2.7% to 8.0%). There were no significant differences in adverse events between the 2 groups. Most physicians were satisfied with the iNCDSS for its clear, time-saving, effective, and safe clinical support.

Conclusions and relevance: In this RCT of an iNCDSS, the system demonstrated noninferiority to senior endocrinology physicians in insulin titration in an inpatient setting, indicating its potential as a favorable tool for insulin titration in patients with T2D.

Trial registration: ClinicalTrials.gov Identifier: NCT04642378.

Figure 1.. CONSORT Diagram

CGM indicates continuous glucose monitoring; iNCDSS, insulin clinical decision support system. ^aThe iNCDSS group had 74 participants with the primary outcome. ^bThe physician group had 70 participants with the primary outcome.

Figure 2.. Sensor Glucose Concentration During Intervention

A, Median sensor glucose concentrations. Solid orange line (orange shaded area) indicate median (IQR) sensor glucose concentration during use of the insulin clinical decision support system (iNCDSS); solid blue line (blue shaded area) indicate median (IQR) number of physician interventions. The brown shaded area indicates overlapping IQRs. The lower and upper limits of the target glucose concentration range of 70 to 180 mg/dL are denoted by the horizontal dashed lines. B, Mean sensor glucose concentration over time among participants who received iNCDSS (orange line) and senior physician (blue line) insulin recommendations. Error bars indicate standard error of the mean. To convert glucose to millimoles per liter, multiply by 0.0555.

Figure 3.. Adverse Events and Physicians’ Satisfaction

A, Proportion of patients with hyperglycemic (>360 mg/dL) or hypoglycemic (<54 mg/dL) episodes. The number of patients with adverse events are shown below the x-axis. There were no statistically significant differences between the insulin clinical decision support system (iNCDSS) group and the physician group. B, Mean scores on the postintervention evaluation of the iNCDSS during the treatment trial assessed by physicians (n = 10) using questionnaires (eMethods in Supplement 2). The satisfaction agreement was scored on a scale of 1 to 5, with 1 indicating very dissatisfied or disagree and 5 indicating very satisfied or agree. Error bars indicate SDs. AI indicates artificial intelligence. To convert glucose to millimoles per liter, multiply by 0.0555.