Truncating Time-to-Event Data for Methodological Investigation

OBJECTIVES: Publications that report early Kaplan Meier (KM) from a clinical trial are useful for understanding how accurate trends seen in immature data are to that of end-of-trial publications. This immature data may not be reported or available. This de novo approach aimed to recreate early trial data from full published KM. If successful, this approach may be used to adjust trial evidence to reflect an earlier timepoint, allowing for before and after comparisons so that analysts can observe predictive data trends and characteristics.

METHODS: Published five-year KM from the ZUMA-1 trial for axicabtagene ciloleucel CAR-T therapy in refractory large B-cell lymphoma was digitised and pseudo individual participant data (IPD) was generated using the Guyot algorithm. The aim was to produce IPD that reflected 12-month, 18-month and 24-month minimum follow-ups, with the proportion censored at a given time informed by earlier ZUMA-1 trial publications.

The censoring indicator was changed until the desired proportion censored was reached, starting with the latest reported events in the five-year dataset. A truncated normal distribution was assumed when assigning the relevant follow-up time for the newly censored points. Generated early datasets were plotted and summary statistics produced, which were compared to equivalent outputs of early ZUMA-1 publications.

RESULTS: The generated truncated datasets resulted in KM that visually matched the KM reported in early ZUMA-1 publications. Median survival was not met in either dataset. The total number of events is 49 in the generated truncated 24-months minimum follow-up dataset and 50 in the ZUMA-1 early publication with 27-month median follow-up.

CONCLUSIONS: Publications reporting trial KM can be truncated to recreate earlier datasets to a reasonably accurate degree. A key feature of the published KM that allowed for this truncation was the small range from minimum to maximum follow-up timepoint, as to narrow the applied censoring to a more accurate range.