Comparison of different machine learning methods and dimensionality reduction for classification astrocytoma and glioblastoma tissues by mass spectra

Abstract

Background: Recently developed methods of ambient ionization allow rapid obtaining of large mass spectrometric datasets, which have a great application in biological and medical analysis. One of the areas that could employ such analysis is neurosurgery. The fast in situ identification of dissected tissues could assist the neurosurgery procedure. The additional information about tumor could help the tumor border monitoring. In this paper, tumor tissues of astrocytoma and glioblastoma are compared, as their identifications during surgery could influence the extent of resection and, hence, the median and overall survival. Methods: Mass spectrometric profiles of brain tumor tissues contain molecular information, which is rather hard to interpret in terms of identifications of individual molecules. The machine learning algorithms are employed for the fast automated mass spectra classification. Different algorithms of dimensionality reduction are considered to process the mass spectra before the classification task, as the initial dimensionality of mass spectra is too high compared with the number of mass spectra. Results: Different classifiers are compared for both just preprocessed data and after dimensionality reduction. The Non-Negative Matrix Factorization appears to be the most effective dimensionality reduction algorithm. The random forest algorithm demonstrated the most robust appearance on the tested data. Also, the comparison of the accuracy of the trained classifier on the mass spectra of tissues measured with different instruments and different resolution is provided in the paper. Conclusions: Machine learning classifiers overfit the raw mass spectrometric data. The dimensionality reduction allows the classification of both train and test data with 88% accuracy. Positive mode data provides better accuracy. A combination of principal component analysis and AdaBoost algorithms appears to be most robust to changing the instrument and conditions.