An Evaluation and Comparison of Methods Used in Survival Analysis To Fit Distributional Curves To Kaplan-Meier Data.

OBJECTIVES: HTA bodies increasingly require accurate survival estimates to provide reliable recommendations. It is argued that access to individual patient data (IPD) can improve their accuracy. This poster aims to assess to what degree extracting IPD from published Kaplan-Meier (KM) curves helps improve extrapolated survival estimates. Some methods currently used for HTA submissions fit a survival curve directly to a published KM curve, but does this lack accuracy? METHODS: Two methods used to extract the IPD from KM curves reviewed in this poster were proposed by Guyot et al[3] and Hoyle and Henley[5] which are compared against the standard least squares method. Comparisons are made for two scenarios: 1) when numbers at risk are available at different time points throughout the KM curve and 2) when numbers at risk are only available at the start. RESULTS: The three methods resulted in the log-normal distribution providing the best fit, with all containing the true mean and median within their confidence intervals (CIs). However, the Hoyle and Henley method estimates a mean marginally closer to the true mean than the other methods in both scenarios. When many numbers at risk are provided, the Hoyle and Henley method gives narrower CIs. Both extraction methods slightly outperform the least squares method. The three methods give median estimates and resulting CIs that are statistically equivalent to that of the IPD, except for the Guyot method when numbers at risk are not available. CONCLUSIONS: Extraction methods give marginally better results than the Least Squares method. However, these results may not be applicable to other case studies. In addition, the extra time taken to run extraction methods may not account for the small improvement in accuracy of results.