Prediction of Pathological and Radiological Nature of Glioma by Mass Spectrometry Combined With Machine Learning

Abstract BACKGROUND: We have previously developed a medical diagnostic pipeline that employs mass spectrometry and machine learning. It does not annotate molecular markers that are specific to cancer but uses entire mass spectra for predicting the properties of glioma. OBJECT: To validate the power of our diagnostic method in predicting the pathological and radiological properties of glioma with a simple sample preparation procedure. METHODS: Ten patients with glioma and 4 non-glioma patients who went through surgical resection were enrolled in our hospital. A total of 1020 mass spectra were acquired from 88 specimens. In order to examine the prediction power of the diagnostic pipeline that we have developed, we performed ten-fold cross-validation for pathological and radiological findings and calculated agreement rates with the conventional methods such as pathological diagnosis (WHO grading, MIB-1 labeling index (LI), mutations in the isocitrate dehydrogenase (IDH)-1 gene and positive 5-ALA fluorescence) and radiological information (gadolinium (Gd)-enhanced area, high-intensity area on fluid-attenuated inversion recovery (FLAIR) imaging,). RESULTS: Prediction accuracy for WHO malignant grade was 91.37%. Those for MIB-1 LI more than 10% and IDH-1 mutation-positive were 82.84% and 87.75%, respectively. Our method achieved an accurate prediction of 95.00% for the 5-ALA-positive lesion. The present method displayed an accuracy of 82.36% in predicting the area of FLAIR hyperintensity and 81.27% for the Gd enhanced area. CONCLUSION: Our methodology achieved a higher rate of prediction of glioma in terms of pathology and radiology. Research is ongoing to develop a validation cohort to verify the biological profiles of glioma specimens.