Random Sample Consensus for the Robust Identification of Outliers in Cancer Data

摘要

Random sample consensus (Ransac) is a technique that has been widely used for modeling data with a large amount of noise. Although successfully employed in areas such as computer vision, extensive testing and applications to clinical data, particularly in oncology, are still lacking. We applied this technique to synthetic and biomedical datasets, publicly available at The Cancer Genome Atlas (TCGA) and the UC Irvine Machine Learning Repository, to identify outliers in the classification of tumor samples. The results obtained by combining Ransac with logistic regression were compared against a baseline classical logistic model. To evaluate the robustness of this method, the original datasets were then perturbed by generating noisy data and by artificially switching the labels. The flagged outlier observations were compared against the misclassifications of the baseline logistic model, along with the evaluation of the overall accuracy of both strategies. Ransac has shown high precision in classifying a subset of core (inlier) observations in the datasets evaluated, while simultaneously identifying the outlier observations, as well as robustness to increasingly perturbed data.