Statistical Analysis and Machine Learning Using Data from Continuous Glucose Monitoring in Clinical Trials

• Activity Number: 653 - Machine Learning and Other Statistical Methods in Clinical Trials
• Type: Contributed
• Date/Time: Thursday, August 1, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Chen Gao* and Yi-Ting Chang and Jay Zhang
• Companies: MedImmune and MedImmune and MedImmune
• Keywords: Continuous Glucose Monitoring; Glucose Control; ANCOVA; Gradient Boosting Machine; Random Forest; Subgroup Analysis

Abstract

Glucose control is important for diabetes management. Using a medical device, Continuous Glucose Monitoring (CGM) captures real-time glucose measurements for a patient every 15 minutes, 24 hours a day. Traditionally, such data are analyzed by patient’s weekly averages, reflecting clinical interpretation similar to Hemoglobin A1c (HbA1c) levels. Ignoring patient’s daily living activities and their impact on glucose level, this approach attributes all deviations from the glucose mean to random variability. CGM was used in a recent randomized Phase II clinical trial to assess the efficacy of a new molecule on glucose control. We used various statistical and machine learning tools to analyze all data points collected during the treatment period to assess important factors accounting for CGM variability. To explore the treatment effects, we first used ANCOVA to analyze the CGM data at several discrete days. We then used gradient boosting machine to demonstrate treatment effect over time. Finally, we employed virtual twins and random forest to investigate important baseline characteristics to explore subgroups with treatment benefit from the new molecule.