Caracterização de dados de sobrevivência em câncer colorretal através de uma abordagem de agrupamento de dados

Abstract

The colorectal cancer is one of the most incident. In the usual clinical practice, survival rates are evaluated from the TNM Staging System, considered the main prognostic system for the cancer. Several works have used different factors and techniques ofmachine learning in search of other factors and prognostic models for several cancer types. The SEER (Surveillance, Epidemiology, and End Results program of the National Cancer Institute) database offers publicly thousands of records of cancer cases in the United States in the last decades, with availability of data for several prognostic factors that are related to the disease. Given the SEER database features, this work has the objective of evaluate the application of some approaches based on data clustering techniques to the survival analysis of colorectal câncer patients: the first is a traditional approach which defines a dissimilaritymatrix and the other an ensemble clustering for obtain the dissimilarity matrix from the evidence accumulation. These techniques are applied in the building of dissimilarities matrices and a prognostic system. Several data clustering algorithms were applied with these approaches, and for each was computed the Akaike Information Criterion (AIC) value. Considering the AIC as criterion for model selection, the results showed that the approach without the use of the ensemble clustering gived better results.