Supervisors info:
Καρλής Δημήτριος, Καθηγητής, Τμήμα Στατιστικής, ΟΠΑ
Μπουρνέτας Απόστολος, Καθηγητής, Τμήμα Μαθηματικώ, ΕΚΠΑ
Δεμίρης Νικόλαος, Επίκουρος Καθηγητής, Τμήμα Στατιστικής, ΟΠΑ
Summary:
The necessity of reliable and valid decision-making regarding public health issues and everyday practice, as well as the continuous accumulation of information in each scientific case, led to the imperative implementation of meta-analyses, aiming at utilizing all the available data and making dependable inferences. Nevertheless, the variety of references in Survival Analysis necessitated the reconstruction of the individual survival data from Kaplan-Meier curves, since they constitute the most commonly published figures, taking into account additional statistical elements too.
In the current thesis, a literature review along with an in-depth explanation of the available methods for reconstructing data from Kaplan-Meier curves and other summary measures at hand have been conducted. In chronological order, these approaches were proposed by Parmar, Williamson, Hoyle & Henley, Guyot, and Rogula. Choosing the most robust method amongst them, i.e, the one suggested by Guyot et al. (2012), and after modifying and enhancing it, a simulation of 50 replications, based on the Weibull distribution and the proportionality assumption, has been carried out. The quality of the reconstructed data has been examined under various alternative scenarios regarding both the number of points extracted from the initial KM graphs and the number of time points at which the size of the risk set is known. Moreover, the consistency between the initial summary measures (HR & RMST) and the ones occurring from the reconstructed data has been investigated.
The results have shown that the larger the number of the extracted coordinates from the initial Kaplan-Meier curves is, the closer to the true value is the algorithm estimate. The discrepancies among the estimates when the number of time points (at which the number of subjects at risk is reported) changes have been negligible. However, it was evident from the Kaplan-Meier curves that the termination of a study of interest occurred earlier in the reconstructed data. Finally, the HR estimates in the applied scenarios were pleasing and robust, especially when additional information was provided. On the contrary, Restricted Mean Survival Time differences estimates exhibited greater variability. Consequently, both the early termination phenomenon in the reconstructed data and the poor estimates of some of the summary measures require further investigation and careful consideration. Additionally, an in-depth examination of the aforementioned method’s accuracy is of the utmost importance on occasions where different distributions are used and under the presence of censoring, two scenarios that reflect better real-life problems.
Lastly, an attempt to convert raster pictures of KM curves to vector images has been made, with the intention of automatizing the coordinate extraction procedure, rectifying in this way the issue of additional errors (OR BIAS) due to manual implementation of the methods. Since this was not possible, it should be stressed that the picture quality, as well as the accuracy of point selection from the investigators, introduce systematic errors, and it is necessary to improve these approaches and optimize them with that in mind.
Keywords:
Survival αnalysis, Kaplan Meier, Individual patient data, Restricted mean survival time, Hazard ratio
