Abstract #0980

Multi-Source Feature Selection to Improve Multi-Class Brain Tumor Typing

Vangelis Metsis¹, Dionyssios Mintzopoulos^2,3, Heng Huang¹, Michael N. Mindrinos⁴, Peter M. Black⁵, Filia Makedon¹, A Aria Tzika^2,3

¹Computer Science, University of Texas, Arlington, TX, USA; ²NMR Surgical Laboratory, MGH & Shriners Hospitals, Harvard Medical School, Boston, MA, USA; ³Radiology, Athinoula A. Martinos Center for Biomedical Imaging, Boston, MA, USA; ⁴Biochemistry, Stanford University School of Medicine, Palo Alto, CA, USA; ⁵Neurosurgery, Brigham and Womens Hospital, Harvard Medical School, Boston, MA, USA

Recent trends in biomedical research have stressed the potential of combining more than one data sources to better understand a patients condition. We acquire state-of-the-art high-resolution magic angle spinning (HRMAS) proton (H1) Magnetic Resonance Spectroscopy (MRS) and gene expression date from the same brain tumor biopsies in order to identify and classify different profiles of brain tumors. We use a novel framework to perform heterogeneous data fusion on both MRS and gene expression datasets using machine learning algorithms. Our experimental results show that our framework outperforms any analysis using individual datasets.