PNS321 EASINESS AND RELIABILITY IN ESTIMATING PATIENT-LEVEL SURVIVAL DATA FROM KAPLAN-MEIER CURVES

Abstract

Studies assessing survival outcomes often do not report sufficient information for inclusion in meta-analysis or decision analytical modeling. Data digitization tools make it possible to extract data from survival curves for the purposes of estimating patient-level survival data and carrying out secondary analysis. Algorithms are available in R and STATA for transforming data. To motivate the use of these methods, this study aimed to demonstrate a) the easiness in extracting data from Kaplan-Meier curves, and b) the accuracy of a published STATA algorithm in estimating patient-level survival data. Data were extracted from Kaplan-Meier curves using the automatic detection function of Webplot Digitizer and then transferred into STATA for estimating patient-level survival data. Figures including different line styles (continuous colored lines, dashed colored lines, black lines with different markers) were selected, within the field of kidney transplantation as an example, and evaluated with and without risk tables. Figures including black lines and different markers were the most difficult to extract data from: It was difficult to discriminate between markers where the curves overlapped and they required the most manual addition and adjustment of data extraction points. A visual inspection of extracted data proved useful in increasing its precision before applying the STATA algorithm. Figures with dashed lines were extrapolated best using the ‘x step with interpolation` option. Estimated survival varied by up to 1% around observed survival; and omitting the risk table did not alter the results. The extraction of survival data from Kaplan-Meier curves is a straightforward process; however, when looking at black-and-white figures the necessity of the desired data should be weighed against the time spent on extraction. The STATA algorithm yielded accurate results across the different scenarios. To capture a more accurate number of events the total number of events may be included in the command.