22-PUB: Machine Learning–Based Optimization of Patient Eligibility for Oral Insulin Treatment—A Post Hoc Analysis of Phase III Trial Results

Background & Objectives: The growing shift toward implementation of precision medicine to improve the prospects of achieving treatment targets requires comprehensive profiling of an array of patient-specific data. Research efforts in this arena include review of anthropometric, epidemiological and clinical features to optimize eligibility criteria for specific treatments. In a recent Phase III, randomized, placebo-controlled, multi-center study in which T2DM patients were treated with oral insulin QD or BID for 26 weeks, the primary endpoint defined as a decrement of 0.6% HbA1c by treatment end was not met. The current post hoc analysis aimed to identify subpopulations with substantial clinical responses to active treatment.

Methods: Data were fitted with a nonparametric causal machine learning model to estimate the impact of various baseline covariates on the expected treatment effect. A recursive partitioning search was then conducted to identify clinically meaningful patient subgroups with high predicted susceptibility to treatment. Results were validated on external datasets, and by running a permutation test and sensitivity analyses to the exact criteria and sample.

Results: Patients with baseline BMI <30.5 kg/m² and age >58 years achieved a conditional average 0.6% reduction in HbA1c by Week 26. Stricter criteria of BMI < 27.0 kg/m² and age >58 years were associated with a mean 0.96% reduction in HbA1c. Gender-specific training of the models suggested different impacts of age and BMI across sex. Male subjects with BMI<30 kg/m² and female subjects > 60 years exhibited improved clinical outcomes.

Conclusions: Variability in response to oral insulin can be minimized by segmenting the patient population based on BMI, age and sex. These findings will inform eligibility criteria for a newly planned Phase III study.