Validation of Insulin Dose Prediction Model Based on Artificial Intelligence Algorithm

Overview

The present study aims to conduct a prospective controlled trial comparing an LSTM-based artificial intelligence (AI) prediction model and clinicians' experience in the efficacy and safety of blood glucose control in hospitalized patients with type 2 diabetes mellitus (T2DM) receiving continuous subcutaneous insulin infusion (CSII) treatment in the Department of Endocrinology. The main question it aims to answer is: Is the prediction model superior to or (at least) non-inferior to clinicians' experience? Eligible patients who receive CSII treatment are randomly allocated into the prediction model group and the empirical group. Patients will: 1. Receive CSII treatment as standard of care during hospitalization for 1-2 weeks, where the daily insulin dose regimen is determined by a prediction model or a clinician's experience. 2. Use continuous glucose monitoring (CGM) for glucose tracking. 3. Receive diabetes self-management education covering nutrition and physical activity.

Data-driven artificial intelligence (AI) represents a new frontier in the modern medical field. The research group previously constructed a database included over 20 years clinical data of short-term intensive insulin therapy (SIIT) via continuous subcutaneous insulin infusion (CSII) in hospitalized patients with type 2 diabetes.The researchers had trained an insulin dose prediction model based on the database using long short-term memory (LSTM) AI algorithms. To establish more robust evidence to validate the efficacy and safety for this model, the present study aims to conduct a pilot prospective controlled trial comparing between AI prediction model and clinician's experience. Specifically, a randomized controlled trial is performed to compare the efficacy and safety of blood glucose control between the insulin dose prediction model and physicians' subjective experience in hospitalized patients with type 2 diabetes mellitus (T2DM) receiving CSII treatment in the Department of Endocrinology. A random number table of 400 participants will be generated using Excel, and randomly allocated into the prediction model group (n=200) and the empirical group (n=200). Medical data will be collected for all the enrolled patients, including medical history, physical examination, auxiliary test reports, continuous glucose monitoring (CGM) data, in which 8 points of the blood glucose (before and 2 hours- postprandial of the 3 main meals, bedtime, and 3 a.m. in the morning) are specifically collected to feed the model for prediction and used for comparisons. For the prediction model group, baseline information upon admission (including age, body mass index \[BMI\], weight, waist circumference, fasting blood glucose before admission and glycated hemoglobin) is put into the model, which will immediately return the insulin dosage (basal rate and boluses for each meal) for the first day of the insulin pump treatment. Physicians will then issue and execute these orders. On the following days, the model adjust the basal rate and boluses based on the previous day's glucose levels and insulin dosages. This process will continue iteratively during the whole CSII period (around 1 to 2 weeks based on whether or not the patient is newly diagnosed or with different disease durations). Insulin pump will be suspended after the administration of dinner bolus on the final day. Fasting blood glucose on the next day after insulin pump suspension will be recorded to conclude the study. In the control (emperical) group, physicians (residents under the guidance of attending doctors) determine insulin dosage based on individual clinical experience and daily glucose monitoring, with patient data collection identical to the prediction model group. Statistical analyses comparing between-group differences in glucose control during CSII treatment, such as time in range (TIR), time below range (TBR), mean blood glucose, glycemic variability, and post-therapy fasting glucose, insulin doses, etc to evaluate the efficacy and saftey of the two insulin dosage determination methods.