Development of a surrogate endpoint for predicting graft failure in kidney transplant patients

GAŠPARAC, GRETA (author), Štrumbelj, Erik (mentor), Aggarwal, Varun (comentor)

Due to the lack of available organs for transplantation it is important to make sure the kidney, once transplanted, survives. Researchers focus on the development of drugs that help prevent graft failure and throughout the process the support provided by statistical models can be beneficial. In this thesis we are given a dataset of patients from three clinics that have undergone kidney transplantation. Our main goal was to develop a model for graft failure prediction at different time points post-transplantation. We analyze the performance of different survival and machine learning models using simulated and real-life data and observe how they compare to the baseline model. Our results indicate that model performance depends on the amount of censored individuals and the amount of individuals that experience graft failure. We advocate for the use of simple models, since more complex approaches, such as joint modelling, often suffer from convergence problems, are more sensitive to model misspecification, and do not seem to add any additional value compared to other approaches. In the thesis we also provide a critical view of related work and review model evaluation metrics. We advocate for the use of inverse probability censoring weighted scoring rules and perform an experiment confirming that they provide unbiased estimates of the metric values under the assumption that censoring is non-informative.

• Language: English
• Keywords: kidney graft failure prediction, medicine, survival analysis, joint modeling
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-141332
• COBISS.SI-ID: 124811267
• Publication date in RUL: 28.09.2022

Secondary language

• Language: Slovenian
• Title: Razvoj nadomestne končne točke za napovedovanje odpovedi presadka pri pacientih s presajeno ledvico

Abstract

Zaradi splošnega pomanjkanja prostih ledvic za presaditev je preživetje presadka ključno. Raziskovalci se osredotočajo na razvoj zdravil za preprečevanje odpovedi presadkov in med znanstvenim procesom so statistični modeli v veliko pomoč. V magistrskem delu imamo na voljo podatke pacientov s presajeno ledvico iz treh različnih študij. Naš cilj je razvoj modela za napovedovanje odpovedi presadka po presaditvi za različne čase. Za lažje razumevanje različnih metodologij in rezultatov ustvarimo tudi umetne podatke. Tako na umetnih kot na realnih podatkih analiziramo različne pristope, od klasičnih modelov preživetvene analize do klasifikacijskih modelov strojnega učenja. Naši rezultati kažejo, da na uspešnost napovedi vplivata odstotek cenzure v učni množici in odstotek dejanskih odpovedi, ki jih zabeležimo med klinično študijo. Zagovarjamo uporabo preprostejših modelov, saj kompleksnejši združeni modeli niso prinesli dodane vrednosti, pogosto pa so imeli težave s konvergenco. V magistrskem delu prav tako podamo kritičen pogled na področje in opozorimo na pomanjkljivosti. Analiziramo metrike ocenjevanja ter zagovarjamo uporabo inverznega verjetnostnega cenzuriranega uteževanja IPCW. Z eksperimentom potrdimo, da so s predpostavko o neinformativni cenzuri metrike IPCW ustrezne in da kljub cenzuri vračajo nepristranske rezultate.

• Keywords: napovedovanje odpovedi presadka ledvic, medicina, preživetvena analiza, skupno modeliranje