All ions must serve: The role of various regimes of data acquisition in joint classifier for intraoperative mass spectrometry-based glial tumour identification

Abstract

Background: Ambient ionisation mass spectrometry, in combination with machine learning techniques, provides a promising tool for rapid intraoperative tumour tissue identification. However, deficiency of non-tumour control samples leads to the classifiers overfitting, especially in neurosurgical applications. Ensemble learning approaches based on the analysis of multimodal mass spectrometry data are able to overcome the overfitting problem through the extended time of data acquisition. In this work, the contribution of each regime of the data acquisition and the requirements for the metrics for further mass spectrometry set-up optimisation are evaluated. Methods: Two independent datasets of the multimodal molecular profiles, a total of 81 glial tumour and non-tumour pathological tissues, were analysed in a cross-validation set-up. The XGboost algorithm was used to build classifiers, and their performance was evaluated using different testing and validating sets. The individual classifiers for each mass spectrometry regime were aggregated into joint classifiers. The impact of each regime was evaluated by the exclusion of specific regimes from the aggregation. Results: The aggregated classifiers with excluded regimes show lower accuracy for most, but not all, excluded regimes. False positive rates have been found to be increased in most cases proving the strong effect of the ensemble learning approach on the overcoming of the “small sample size” problem. Conclusions: The impact of each group of regimes – with different ion polarity, resolution or mass range of spectra was found to be non-linear. It might be attributed to biochemical reasons as well as to the physical limitation of mass analysers. The required metrics for the evaluation of each regime contribution to the classification efficiency should be a numerical estimation of how the classifier depends on any given regime and could not be estimated only by excluding any group of regimes at all.