Date: Sunday, June 12, 2016
Session Name: Poster Session B: Kidney: Cardiovascular and Metabolic
Session Time: 6:00pm-7:00pm
Presentation Time: 6:00pm-7:00pm
Location: Halls C&D
Understanding the risk factors that predispose to death is important in order to deliver the most appropriate therapy to patients (pts) with kidney transplantation (Tx). Aim of this study is to build reliable decision models and to identify the relevant risk factors for death using different data mining approaches.
We analyzed 761 pts transplanted between 2000 and 2007 (follow-up of up to 10 years). Data included biopsy results, clinical and laboratory factors. After feature selection by conventional statistics (28 variables) models were build using Naïve Bayesian, C5.0, RPART and Random Forest.
Compared to C5.0 and RPART, Naïve Bayesian and Random Forest resulted in models with a higher sensitivity to predict death and a high specificity. Using 60% of the data for the training set and 40% for the test set, the model with Naïve Bayesian had a sensitivity of 66% and a specificity of 91% to predict death. With Random Forest, sensitivity was 28% and specificity 98%. Because of the imbalance of the outcome groups (i.e. 13% deceased pts) modeling was repeated with balanced datasets obtained by oversampling the minority class. With the balanced data, sensitivity was 82% and specificity 79% with Naïve Bayesian. With Random Forest, sensitivity was 77% and specificity 88% to predict death. These two models were externally validated with a separate dataset (300 pts), resulting in a sensitivity of 59% and a specificity of 78% for the Naïve Bayesian model and in a sensitivity of 65% and specificity of 82% for the Random Forest model.
Highly important variables were recipient age, high systolic and low diastolic blood pressure, pre-Tx diabetes mellitus, and peripheral arterial and coronary heart disease, annual GFR loss, delayed graft function and time on dialysis. Variables of modest importance included donor age, post-Tx hyperparathyroidism and best graft function within the first 6 weeks and cold ischemia time.
The established models permit reliable prediction of death and survival and can be used to identify patients on risk. Moreover, with the identified (modifiable) risk factors patients can be assigned to different treatment strata to offer each individual the optimal therapy.
CITATION INFORMATION: Scheffner I, Hua K, Simovici D, Abeling T, Haller H, Gwinner W. Prediction of Patient Survival After Kidney Transplantation: Construction, Validation and Evaluation of Decision Models Using Data Mining Approaches. Am J Transplant. 2016;16 (suppl 3).
To cite this abstract in AMA style:Scheffner I, Hua K, Simovici D, Abeling T, Haller H, Gwinner W. Prediction of Patient Survival After Kidney Transplantation: Construction, Validation and Evaluation of Decision Models Using Data Mining Approaches. [abstract]. Am J Transplant. 2016; 16 (suppl 3). https://atcmeetingabstracts.com/abstract/prediction-of-patient-survival-after-kidney-transplantation-construction-validation-and-evaluation-of-decision-models-using-data-mining-approaches/. Accessed March 5, 2021.
« Back to 2016 American Transplant Congress