Supervisors info:
Βασιλική-Αναστασία Σύψα, Αναπληρώτρια Καθηγήτρια, Ιατρική Σχολή, ΕΚΠΑ, Επιβλέπουσα
Παναγιώτα Τουλούμη, Καθηγήτρια, Ιατρική Σχολή, ΕΚΠΑ
Γεώργιος Παπαθεοδωρίδης, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Summary:
The aim of this thesis is the review and the description of methods to develop and validate predictive models. Moreover, the methodology for the development of the risk score under the presence of competing risks was reviewed. After this review a application to available data from patients with chronic hepatitis B was performed in order to predict the risk of developing hepatocellullar carcinoma taking in account the competing events.
We make a reference to basic definitions of survival analysis (survival data, hazard function, censored observations, etc), to Cox proportional hazards model and to various checks of proportional hazards assumptions (graphical checks, through residuals, etc). The competing risks often appear in survival data analysis. We refer to methods of statistical analysis under competing risks (cumulative incidence function, cause-specific hazard function and sub-distribution hazard function) and to regression models with the presence of competing risks, like the semi-parametric model of proportional risks by Fine&Gray for the prediction of sub-distribution hazard.
In order to develop a predictive model, which takes in account the competing risks, it is necessary to define the outcome, all the predictive factors and the competing event. In the presence of competing risks, the appropriate model should be chosen Fine&Gray model. The next steps include evaluation of missing values , coding of predictors, model specification and model estimation. There are methods to estimate the coefficients (LASSO, ridge regression) to reduce overfitting of the model. After the choice of the final mode, a validation (internal or external) and, an evaluation of the calibration, discrimination and its clinical usefulness are performed. In the thesis, the methodology for the development of a risk score with or without competing risks was described.
The application of these methods was based on data from PAGE-B risk score (patients from nine centers with chronical hepatitis B). The outcome was the the development of hepatocellular carcinoma (HCC) in five years after treatment initiation. The competing events were death and liver transplantation. A forward stepwise method was and there were four variables included in the model: age, sex, platelet count and cirrhosis. Internal validation was performed using bootstrap and the resulting c-index was 0.081 based on this score, a points-based risk score was developed to predict the 5-year risk of HCC development. Patients with score ≤9 had low risk of HCC, whereas patients with score ≥18 had high 5-year risk. The difference between the risk score taking into account competing events and PAGE-B were negligible due the low number of competing events in this dataset.
Prediction models are now used widely and they provide an important tool for clinicians. As these models may guide decisions related to the follow-up and treatment if patients it is important to strengthen methodological rigor. Competing risks should be consider in the development of prediction models when analyzing survival data.
