A comparative study of data mining techniques applied to renal-cell carcinomas

Abstract

Despite being one of the deadliest diseases and the enormous evolution in fighting it, the best methods to predict kidney cancer, namely Renal-Cell Carcinomas (RCC), are not well-known. One of the solutions to accelerate the current knowledge about RCC is through the use of Data Mining techniques based on patients' personal and clinical data. Therefore, it is crucial to understand which techniques are the most suitable to extract knowledge about this disease. In this paper, we followed the CRISP-DM methodology to simulate different techniques to determine the ones with the best predictive performance. For this purpose, we used a dataset of 821 records of RCC patients, obtained from The Cancer Genome Atlas. The present work tests different Data Mining techniques, that can be used to predict the 5-year life expectancy of patients with renal cancer and to predict the number of days to death for patients who have a life expectancy of less than 5 years. The results obtained demonstrated that the best algorithm for estimating the vital status at 5 years was Random Forest. This algorithm presented an accuracy of 87.65% and an AUROC of 0.931. For the prediction of days to death, the best performance was obtained with the k-Nearest Neighbors algorithm with a root mean square error of 354.6 days. The work suggested that Data Mining techniques can help to understand the influence of various risk factors on the life expectancy of patients with RCC.