Outlier detection and ambiguity detection for microarray data in probabilistic discriminant partial least squares regression

The reject option plays an important role in the classification of microarray data. In this work, a reject option is implemented in the probabilistic discriminant partial least squares (p-DPLS) method in order to reject to classify both outliers and ambiguous samples. Microarray data are highly susceptible to present outliers because of the many steps involved in the experimental process. During the development of the classifier, outliers in the training data may strongly influence the model and degrade its performance. Some future samples to be classified may also be outliers that will most probably be misclassified. Ambiguous samples are samples that cannot be clearly assigned to any of the classes with a high confidence. In this work, outlier detection and ambiguity detection are implemented taking into account the x-residuals, the leverage and the predicted (y) over cap. The method was applied to oligonucleotide microarray data and cDNA microarray data. For the first dataset (prostate cancer data set), the outlier detection criteria allowed us to remove nine samples from the training set. The model without those samples had better classification ability, with a decrease in the classification cost per sample from 0.10 to 0.07. The method was also used in a second dataset (small round blue cell tumours of childhood dataset) to detect prediction outliers so that most of the outliers were rejected to classify and misclassifications were reduced from 100 to 5%. Copyright (C) 2010 John Wiley & Sons, Ltd.