COMPSTAT 2016: Start Registration
View Submission - COMPSTAT
Title: Regularised multiblock methods for cancer patient classification Authors:  Tommy Lofstedt - Umea University (Sweden) [presenting]
Patrik Brynolfsson - Umea University (Sweden)
Thomas Asklund - Umea University (Sweden)
Tufve Nyholm - Umea University (Sweden)
Abstract: Regularised generalised canonical correlation analysis (RGCCA) was recently extended to allow for the connections between blocks to go beyond the covariance link, and to predict a single binary outcome variable through logistic regression. The generalisation is called multiblock logistic regression (Multiblog) and contains logit links between the blocks and the outcome, and covariance links between the blocks. We extend the Multiblog model by allowing the path coefficients to be $c_{ij}\geq 0$, to control the influence of the covariance on the logistic regression models, and by regularising the regression coefficients. The aim was to study gray-level co-occurrence matrices (GLCMs) computed from three types of MRI images from 23 patients with high-grade glioma. The MRI images were the estimate of the apparent diffusion coefficient (ADC), T1- and T2-weighted images. These three types constituted three blocks, and the purpose was to predict whether the tumour would regress or progress, or whether the patient would exceed the median survival time or not. We compared the performance of these new extended RGCCA models and of several classical machine learning and multiblock methods, with or without regularisation, when applied on the GLCM data to those of the traditional Haralick feature extraction approach.