Enrico Glaab1, Jean-Pierre Trezzi2, Andrea Greuel3, Christian Jäger4, Zdenka Hodak4, Alexander Drzezga5, Lars Timmermann3, Marc Tittgemeyer6, Nico Jean Diederich7, Carsten Eggers3. 1. Luxembourg Centre for Systems Biomedicine, Esch-sur-Alzette, Luxembourg. Electronic address: enrico.glaab@uni.lu. 2. Luxembourg Centre for Systems Biomedicine, Esch-sur-Alzette, Luxembourg; Integrated Biobank of Luxembourg, Luxembourg Institute of Health, Dudelange, Luxembourg. 3. Clinic for Neurology, University Hospital Giessen and Marburg, Marburg, Germany. 4. Luxembourg Centre for Systems Biomedicine, Esch-sur-Alzette, Luxembourg. 5. Department of Nuclear Medicine, University Hospital Cologne, Cologne, Germany. 6. Max Planck Institute for Metabolism Research, Cologne, Germany; Cologne Cluster of Excellence in Cellular Stress and Aging-Associated Disease (CECAD), Cologne, Germany. 7. Department of Neurology, Centre Hospitalier du Luxembourg, Luxembourg-City, Luxembourg.
Abstract
BACKGROUND: The diagnosis of Parkinson's disease (PD) often remains a clinical challenge. Molecular neuroimaging can facilitate the diagnostic process. The diagnostic potential of metabolomic signatures has recently been recognized. METHODS: We investigated whether the joint data analysis of blood metabolomics and PET imaging by machine learning provides enhanced diagnostic discrimination and gives further pathophysiological insights. Blood plasma samples were collected from 60 PD patients and 15 age- and gender-matched healthy controls. We determined metabolomic profiles by gas chromatography coupled to mass spectrometry (GC-MS). In the same cohort and at the same time we performed FDOPA PET in 44 patients and 14 controls and FDG PET in 51 patients and 16 controls. 18 PD patients were available for a follow-up exam after one year. Both data sets were analysed by two machine learning approaches, applying either linear support vector machines or random forests within a leave-one-out cross-validation scheme and computing receiver operating characteristic (ROC) curves. RESULTS: In the metabolomics data, the baseline comparison between cases and controls as well as the follow-up assessment of patients pointed to metabolite changes associated with oxidative stress and inflammation. For the FDOPA and FDG PET data, the diagnostic predictive performance (DPP) in the ROC analyses was highest when combining imaging features with metabolomics data (ROC AUC for best FDOPA + metabolomics model: 0.98; AUC for best FDG + metabolomics model: 0.91). DPP was lower when using only PET attributes or only metabolomics signatures. CONCLUSION: Integrating blood metabolomics data combined with PET data considerably enhances the diagnostic discrimination power. Metabolomic signatures also indicate interesting disease-inherent changes in cellular processes, including oxidative stress response and inflammation.
BACKGROUND: The diagnosis of Parkinson's disease (PD) often remains a clinical challenge. Molecular neuroimaging can facilitate the diagnostic process. The diagnostic potential of metabolomic signatures has recently been recognized. METHODS: We investigated whether the joint data analysis of blood metabolomics and PET imaging by machine learning provides enhanced diagnostic discrimination and gives further pathophysiological insights. Blood plasma samples were collected from 60 PDpatients and 15 age- and gender-matched healthy controls. We determined metabolomic profiles by gas chromatography coupled to mass spectrometry (GC-MS). In the same cohort and at the same time we performed FDOPA PET in 44 patients and 14 controls and FDG PET in 51 patients and 16 controls. 18 PDpatients were available for a follow-up exam after one year. Both data sets were analysed by two machine learning approaches, applying either linear support vector machines or random forests within a leave-one-out cross-validation scheme and computing receiver operating characteristic (ROC) curves. RESULTS: In the metabolomics data, the baseline comparison between cases and controls as well as the follow-up assessment of patients pointed to metabolite changes associated with oxidative stress and inflammation. For the FDOPA and FDG PET data, the diagnostic predictive performance (DPP) in the ROC analyses was highest when combining imaging features with metabolomics data (ROC AUC for best FDOPA + metabolomics model: 0.98; AUC for best FDG + metabolomics model: 0.91). DPP was lower when using only PET attributes or only metabolomics signatures. CONCLUSION: Integrating blood metabolomics data combined with PET data considerably enhances the diagnostic discrimination power. Metabolomic signatures also indicate interesting disease-inherent changes in cellular processes, including oxidative stress response and inflammation.
Authors: Marina C Ruppert; Andrea Greuel; Julia Freigang; Masoud Tahmasian; Franziska Maier; Jochen Hammes; Thilo van Eimeren; Lars Timmermann; Marc Tittgemeyer; Alexander Drzezga; Carsten Eggers Journal: Hum Brain Mapp Date: 2021-02-27 Impact factor: 5.038