OBJECTIVES: Survival data of IO therapies often contain inflection points indicating changes in hazard at different time points. Standard parametric regression models may not fit such data well, with piecewise fitting potentially providing a more robust estimate of survival projection. However, long-term survival estimates are highly sensitive to the choice of cut-off points and generally no formal analyses are conducted for cut-off point selection. The aim of this study was to use statistical tests for optimal cut-off point selection in piecewise models.

METHODS: Ten overall survival (OS) curves of IO therapies in different indications were digitised and the individual patient-level data recreated. Kaplan-Meier (KM) estimates were converted to log-cumulative hazards and ordinary-least square (OLS) regression were fitted to the data. To test the null hypothesis, i.e. ‘no structural change’ present in the data, generalised fluctuation tests and F test statistics were used along with a boundary indicating the 5% critical value for the test statistic. From the F statistics, optimal inflection points were determined via Bayesian Information Criterion.

RESULTS: Inflection points were identified for all OS curves. In cases where OLS regression applied to the log cumulative hazard data showed an R² as high as 0.93 (considered a good fit), an optimal point of inflection was identified using the above algorithm. Furthermore, parametric curves were fitted to KM data at the optimal inflection point and the resulting model fit was substantially improved compared with previous analyses where full parametric curves were used.

CONCLUSIONS: Studies in the literature indicate that a delay in IO treatment response exists prior to the inflection point, after which durable treatment benefit may be observed. This is aligned with our findings. The method used in this study provides a framework for the identification inflection points which will provide robust long-term survival estimates.