Applying the inverse probability of censoring weighting approach in survival analysis based on machine learning methods

Abstract

Introduction: Machine learning methods for right censored survival data lead to biased estimates or less accurate risk predictions. Objectives: The purpose of this study was to apply the inverse probability of censoring weighting (IPCW) approach in machine learning techniques including decision tree (DT), k-nearest neighbors (KNN) and generalized additive model (GAM) to provide better estimates of censored times and more accurate predictions for breast cancer data. Methods: We used data of 1154 newly diagnosed breast cancer (BC) cases recorded in the East Azerbaijan population-based cancer registry database between March 2007 and March 2016. Three machine learning techniques approach with IPCW technique was used to assess the association between mortality and sex, age, grade, morphology and time. The results of these models were compared using sensitivity, specificity, accuracy, area under ROC curve, positive predictive value and negative predictive value. Results: A total of 217 (18.8%) individuals experienced death due to BC by the end of the study. Among the fitted models, the GAM had the best fit with sensitivity= 98.8, specificity= 83.9 and accuracy= 92.8. In general, age, grader, morphology and survival time had an effect in correct predictions. Conclusion: The GAM, due to its high predictive power, is recommended for prediction in data of patients with breast cancer. Also, according to significant association of age, grade, morphology and survival time with mortality, considering these factors in the treatment process can be effective in reducing mortality from BC.