Reference : Combining PET imaging and blood metabolomics data to improve machine learning models ...
Scientific congresses, symposiums and conference proceedings : Poster
Life sciences : Biotechnology
Life sciences : Multidisciplinary, general & others
Human health sciences : Neurology
Systems Biomedicine
http://hdl.handle.net/10993/37568
Combining PET imaging and blood metabolomics data to improve machine learning models for Parkinson’s disease diagnosis
English
Glaab, Enrico mailto [University of Luxembourg > Luxembourg Centre for Systems Biomedicine (LCSB) > >]
Trezzi, Jean-Pierre [University of Luxembourg > Luxembourg Centre for Systems Biomedicine (LCSB) > >]
Greuel []
Jäger, Christian [University of Luxembourg > Luxembourg Centre for Systems Biomedicine (LCSB) > >]
Hodak, Zdenka [University of Luxembourg > Luxembourg Centre for Systems Biomedicine (LCSB) > >]
Timmermann, Lars []
Tittgemeyer, Marc []
Diederich, Nico Jean []
Eggers, Carsten []
8-Oct-2018
Yes
International
2018 International Congress of the International Parkinson and Movement Disorders Society
from 05-10-2018 to 09-10-2018
Esch-sur-Alzette
Luxemburg
[en] Parkinson ; neuroimaging ; metabolomics ; PET ; machine learning ; integration ; differential abundance
[en] Objective: To investigate whether the integration of PET imaging and metabolomics data can provide improved machine learning models for PD diagnosis.

Background: The reliable diagnosis of PD can remain challenging, even at the motor stage. PET imaging can be used to confirm the clinical diagnosis. However, limitations in the robustness of predictive features extracted from the data and the costs associated with PET imaging restrict its application. Using blood metabolomics data as an additional information source may provide improved combined diagnostic models and/or an initial filter to decide on whether to apply PET imaging.

Methods: Metabolomics profiling of blood plasma samples using gas chromatography coupled to mass spectrometry (GC­MS) was conducted in 60 IPD patients and 15 healthy controls. After pre-processing, these data were compared to neuroimaging data for subsets of the same individuals using FDOPA PET (44 patients and 14 controls) and FDG PET (51 patients and 15 controls). Machine learning models using linear support vector machines were trained on 50% of the data and evaluated on a 50% hold­out test set using Receiver Operating Characteristic (ROC) curves. Next, standardized FDOPA and FDG PET intensity measurements were combined with those from the metabolomics data to build and evaluate sample classification models in the same manner as for the individual datasets.

Results: Both for the FDOPA and FDG PET data, the predictive performance given by the area under the ROC curve (AUC) was highest when combining imaging features with those from the metabolomics data (AUC for FDOPA + metabolomics: 0.98; AUC for FDG + metabolomics: 0.91). The performance was generally lower when using only the respective PET attributes (FDOPA: 0.94, FDG: 0.8) or only the metabolomics data (AUC: 0.66).
Luxembourg Centre for Systems Biomedicine (LCSB): Biomedical Data Science (Glaab Group)
Fonds National de la Recherche - FnR
Researchers ; Professionals ; Students
http://hdl.handle.net/10993/37568
FnR ; FNR7643563 > Rudi Balling > Mito-PD > Mitochondrial endophenotypes of PD > 01/01/2015 > 31/12/2017 > 2013

File(s) associated to this reference

Fulltext file(s):

FileCommentaryVersionSizeAccess
Open access
metabolomics PET machine learning poster.pdfAuthor postprint327.37 kBView/Open

Bookmark and Share SFX Query

All documents in ORBilu are protected by a user license.