Premium
Individualized MR‐based prediction of cognitive performance in subjects at risk of dementia
Author(s) -
Nemali Aditya Sai Ram,
Yakupov Renat,
Schütze Hartmut,
Spottke Annika,
Ramirez Alfredo,
Schneider Anja,
Metzger Coraline D.,
Christoph Laske,
Bittner Daniel,
Brosseron Frederic,
Priller Josef,
Wiltfang Jens,
Buerger Katharina,
Fließbach Klaus,
Heneka Michael T.,
Peters Oliver,
Speck Oliver,
Nestor Peter J.,
Teipel Stefan J.,
Pross Verena,
Glanz Wenzel,
Wagner Michael,
Jessen Frank,
Düzel Emrah,
Ziegler Gabriel
Publication year - 2021
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.053018
Subject(s) - dementia , neuroimaging , biomarker , neuropsychology , voxel based morphometry , cohort , alzheimer's disease neuroimaging initiative , neurocognitive , psychology , disease , voxel , atrophy , cognition , white matter , medicine , artificial intelligence , machine learning , magnetic resonance imaging , neuroscience , computer science , radiology , biology , biochemistry
Background Neuroimaging markers based on MRI often provide better prediction than traditional neuropsychological scores. With advancements of machine learning, data patterns may offer opportunities to personalize clinical practice that leads to better outcomes for patients at risk of dementia such as Alzheimer’s disease (AD) (Davatzikos et al., 2019). AD is a multifactorial process associated with ageing, brain atrophy, genes, proteins, vascular risk, and brain state activity (Frisoni et al., 2010). These processes do covary and interact in a complex fashion which needs to be accounted when aiming at predicting clinical outcomes for staging and stratification of disease‐modifying treatments. Method In our probabilistic predictive framework we focus on data from the DZNE DELCODE cohort (Jessen et al., 2018) consisting of T1‐weighted and FLAIR images to assess distributed patterns of Voxel‐based Morphometry (VBM) and White Matter Lesions for 929 subjects; subject‐specific demographics (age, sex, education) and available CSF biomarkers for 438 subjects. We developed a machine learning framework for brain‐based predictions of (A) memory performance (Wolfsgruber et al., 2020) and (B) CSF Amyloid 42/40 and p‐tau biomarker status using a Gaussian process multi‐kernel (GP‐MKL) learning approach (Rasmussen & Williams, 2006). The proposed GP‐MKL model combines multiple features (atrophy patterns, demographics age, sex, education, white matter lesions volume & apoe4) expected to characterize the transition from healthy ageing towards dementia in terms of cognitive symptoms and biomarker status (Figure 1). We evaluate predictive models and different feature combinations using 10‐fold cross‐validation. Result The framework enabled optimal individual prediction of memory performance (highest correlation true vs. predicted of r = 0.751 ± 0.082, R 2 = 0.56, Fig. 2) using a combination of demographics, brain tissue segments (GM & CSF) & CSF biomarkers (Aß42/40 & p‐tau). When estimating the CSF biomarker positivity, the AUC‐ROC score achieved 0.735 for Aß42/40 (Fig. 4A) and 0.802 for p‐tau (Fig. 4B) using a combination of brain tissue segments (GM & CSF), demographics, and cognitive testing. Conclusion In conclusion, multiple domains and imaging facets contribute to reliable estimation of individual cognitive memory performance and biomarker positivity in dementia and enable promising predictive technologies for staging and treatment stratification.