Prediction-driven decision rules, RCT design and survival analysis

Predictions are becoming more and more a part of our lives, and they are becoming increasingly useful in medical science as the science evolves. Increased understanding of disease and its treatments allows us to use predictions based on predictive biomarker signatures to optimize treatment outcomes for increasingly granular subject groups. One such potential use is in the field of HIV treatment monitoring. In resource-limited regions where regular testing for HIV treatment failure is not always possible, pooled testing methods can reduce the burden of regular testing for all infected. Incorporating predictions to choose who is individually tested based on pooled test results is a way to increase the efficiency of such methods, the treatment being the individual testing versus pooled testing only.

The use of biomarker-guided treatment decision rules, or prediction-driven decision rules, can be informal or formally well-defined. For a well-defined prediction-driven decision rule to be implemented, it must first be rigorously tested for efficacy based on a comparison against the standard of care. The definition of standard of care and thus, the definition of clinical utility, depends heavily on the treatment setting. Poorly defining clinical utility can result in great bias, potentially leading to implementing unnecessary prediction-driven decision rules.

Formal prediction-driven decision rules are currently most applied in the disease area of cancer. Rigorous testing of these rules is often conducted through RCTs, specifically group sequential RCTs, utilizing a survival endpoint. It is important to understand the analysis of survival data in order to ensure the appropriate analysis methods for such data. Confidence bands for survival estimates over time should be constructed to have nominal coverage rates, and analysis methods like RMST should be understood to allow for rigorous testing of differences when proportional hazards assumptions are not met.

Developing prediction-driven decision rules in the form of pooled testing methods for HIV treatment failure, identifying an RCT trial design(s) capable of rigorously evaluating these prediction-driven decision rules, and studying survival analysis methods capable of analyzing the data from such RCTs, whether proportional hazards holds or not, are the subjects of this dissertation.

List of scientific papers

I. Brand A, May S, Hughes JP, Nakigozi G, Reynolds SJ, Gabriel EE. Prediction‐driven pooled testing methods: Application to HIV treatment monitoring in Rakai, Uganda. Statistics in Medicine. 2021 Aug 30;40(19):4185-99.
https://doi.org/10.1002/sim.9022

II. Brand A, Sachs MC, Sjölander A, Gabriel EE. Confirmatory prediction-driven RCTs in comparative effectiveness settings for cancer treatment. British Journal of Cancer. 2023 Jan 23:1-8.
https://doi.org/10.1038/s41416-023-02144-x

III. Sachs MC, Brand A, Gabriel EE. Confidence bands in survival analysis. British Journal of Cancer. 2022 Nov 1;127(9):1636-41.
https://doi.org/10.1038/s41416-022-01920-5

IV. Brand A, Sachs MC, Gabriel EE. Estimating differences in restricted mean survival time in R with two new implementations. [Manuscript]

V. Brand A, Sachs MC, Gabriel EE. Evaluating restricted mean survival time methods in group sequential RCTs. [Manuscript]