Background: In randomized controlled trials (RCTs), assay sensitivity issues can affect both statistical power and the confidence in treatment efficacy estimates. This sensitivity problem is often linked to the role of contextual effects in patient response. While statistical adjustments for prognostic variables can address this, it is impractical to account for all possible covariates. Machine learning-based models, which combine many covariates into a single prognostic index, offer a solution. We evaluated the transferability of the Placebell model, originally tested on historical data from other conditions (e.g., pain, Parkinson’s disease), to Type 1 Diabetes (T1D) studies.

Methods: The Placebell model was calibrated using data from chronic pain RCTs and designed to integrate baseline factors like disease severity, psychological factors, and demographics. To assess its applicability in T1D, comparable data were used to calculate prognostic scores in the IMPACT study and an associated substudy with 85 and 24 patients, respectively. These prognostic scores were used to adjust the analysis of four endpoints (two C-peptide responses, average insulin consumption, and HbA1c levels).

Results: The model consistently enhanced analysis precision, with improvements ranging from 2.9% to 52.2% for C-peptide responses and 1.6% to 20.9% for other endpoints. This increase in precision is equivalent to an effective sample size expansion by the same proportions, resulting in a gain of up to 44 additional patients in the main study.Conclusion