Supervisors info:
Σύψα Βασιλική-Αναστασία, Επίκουρη Καθηγήτρια, Ιατρική, ΕΚΠΑ
Μπουρνέτας Απόστολος, Καθηγητής, Τμήμα Μαθηματικών, ΕΚΠΑ
Μπάμια Χριστίνα, Αναπληρώτρια Καθηγήτρια, Ιατρική, ΕΚΠΑ
Summary:
This thesis aims to provide a review and a detailed description of the methods used to develop and validate prognostic models for binary outcomes; to develop a review of the method by which a risk score can be constructed from such a prognostic model, to facilitate its use in everyday practice as well as to apply these methods on real world data.
Determining the outcome and all candidate prognostic factors, while choosing the right model according to the data are the initial steps required to form a prognostic model. The logistic regression model is appropriate for binary outcomes. As a next step, a proper management of missing values according to their type is necessary. The coding of continuous variables comes next; it can be done in various methods depending on the variable (Dichotomization, Categorization, Linear, Polynomials, Fractional polynomials and Splines). The method of selecting prognostic risk factors is then selected (e.g. stepwise methods or bootstrap). In addition, the final model could include all possible risk factors, without any further reduction, through a selection procedure. Once the model selection method is chosen, the regression coefficients are evaluated, usually through the Maximum Likelihood Method. However, new methods are reported in the literature (Uniform Shrinkage Method, Penalized ML Method, LASSO Method, Ridge Method) to limit model overestimation.
It is necessary to validate the final model with internal or external validation and to assess its calibration ability, its descrimination ability, as well as its clinical utility. Finally, creating a risk score for the model contributes to facilitating its use in clinical practice.
These methods are applied on data from the Athenian General Hospital “Laiko” in order to assess the risk factors for MRSA colonization on admission of patients. A prognostic tool for identifying these patients would be useful to simplify the MRSA control policy in hospitals. The application of a forward stepwise selection identified seven statistically significant risk factors for MRSA infection: diabetes, nursing home resident, recent use of antibiotics, dementia / psychological disorders, chronic skin disease, recent hospitalization, gender. Consiquently, internal validation in the model were obtained: c-statistic equal to 1.031 with 95% CI (0.891, 1.171), AUC equal to 0.768 with 95% CI (0.697.0.839) and discrimination slope equal to 0.051.
The final prognostic model based on the internal validation applied, appears to satisfactorily predict the outcome. However, external validation would provide more information about the validity of this prognostic model.
Modern medicine is increasingly based on diagnostic prognostic models to inform individuals and health professionals about the risks of occurrence of an outcome and to guide clinical decisions designed to mitigate these risks. Therefore, prognostic models are becoming more and more useful in medical practice.
Keywords:
prognostic models, binary outcomes, validation of prognostic model, development of prognostic model
