OBJECTIVES: All indirect treatment comparisons (ITC) rely on individual patient-level data (IPD) to compare the efficacy of healthcare interventions. However, IPD are often only available for the index study while Kaplan-Meier (KM) and aggregate data are used for the comparator trials. Techniques exist to generate pseudo-IPD from KM outcome data, but this only allows estimation of the outcome and not the covariates. This research aimed to develop a method to generate pseudo-IPD by predicting the covariates on an individual patient level for use in ITCs.

METHODS: The “2-opt swap & shuffling” algorithm uses KM estimates to get the time and event per patient. Subgroup information is used for the hazard ratio (HR) and number of patients. In each iteration, patients are either swapped or shuffled, and an absolute error is calculated. If the new error is lower, this becomes the new solution. A base case was tested to evaluate the general performance of the algorithm and explore the effects of the following: a) when a subgroup has an (in)significant HR; b) imbalances between the number of patients per subgroup; c) different sample sizes; and d) when all desired information is not available (e.g., no median survival per subgroup). Five datasets (using KM data from an existing study with two treatments) were run with a cutoff time of 10 minutes.

RESULTS: The algorithm produced an almost perfect replication of the KM per subgroup in the base-case scenario. The best performance was on the semi-significant HR data, with minimal errors in the five datasets.

CONCLUSIONS: This novel algorithm can recreate subgroup KM data for different covariates with minimal errors. The method has the potential for broader use in ITCs, such as effect-modifier testing and multivariate regression on IPD.