Comparison of Cox Model and K-Nearest Neighbor to Estimation of Survival in Kidney Transplant Patients

• Main Authors: Javad Faradmal, Tahereh Omidi, Jalal Pourolajal, Ghodratollah Roshanaei
• Format: Article
• Language: fas
• Published: Hamadan University of Medical Sciences 2016-03-01
• Series: پزشکی بالینی ابن سینا
• Subjects: cox model; kidney transplantation; k- nearest neighbor; transplantation rejection
• Online Access: http://sjh.umsha.ac.ir/article-1-557-en.html

Introduction & Objective: Cox model is a common method to estimate survival and validity of the results is dependent on the proportional hazards assumption. K- Nearest neighbor is a nonparametric method for survival probability in heterogeneous communities. The purpose of this study was to compare the performance of k- nearest neighbor method (K-NN) with Cox model. Materials & Methods: This retrospective cohort study was conducted in Hamadan Province, on 475 patients who had undergone kidney transplantation from 1994 to 2011. Data were extracted from patients’ medical records using a checklist. The duration of the  time between kidney transplantation and rejection was considered as the surviv­al time. Cox model and k- nearest neighbor method were used for Data modeling.  The prediction error Brier score was used to compare the performance models. Results:  Out of 475 transplantations, 55 episodes of rejection occurred. 5, 10 and 15 year survival rates of transplantation were 91.70 %, 84.90% and 74.50%, respectively. The number of neighborhood optimized using cross validation method was 45. Cumulative Brier score of k-NN algorithm for t=5, 10 and 15 years were 0.003, 0.006 and 0.007, respectively. Cumulative Brier of score Cox model for t=5, 10 and 15 years were 0.036, 0.058 and 0.058, respectively.  Prediction error of k-NN algorithm for t=5, 10 and 15 years was less than Cox model that shows that the k-NN method outperforms. Conclusions: The results of this study show that the predictions of KNN has higher accuracy than the Cox model when sample sizes and the number of predictor variables are high.