Diagnostic modelling of multimodal neuroimaging data in mental disorders
Nuclear Medicine, Radiotherapy, Radiobiology
Final Report Abstract
Advanced multimodal modelling of neuroimaging data in mental disorders has delivered additional insights into disease mechanisms. Only initial experience has been gained in applying those techniques for diagnostic modelling. Purposes of this project were: a) to implement an explicitly diagnostic modelling approach for joint analyses of multimodal MRI datasets in mental disorders that can optionally be extended by further sources of information, b) to optimize and validate this approach in existing large-scale datasets (major depressive disorder, ageing with a focus on dementia) and c) to arrange its application to further diagnostic questions when sufficient diagnostic accuracies can be achieved. A multimodal modelling approach was implemented which combined advanced singlemodality data processing techniques based on well-established approaches with joint multimodal modelling using the recently developed linked independent component analysis. Resulting multimodal representations (components consisting of spatial maps, modality weights and individual subject weights) were assessed comprehensively for their relationship with relevant disease measures. Subject weights, representing the strength and directionality of the relationship of an individual subject with a multimodal component, were used as features for diagnostic modelling. These models mainly used machine learning techniques. This approach was tested with appropriate adjustments and automated model selection in a large-scale cross-sectional neuroimaging dataset covering grey matter volume, white matter microstructure and intrinsic functional connectivity in major depressive disorder (MDD) and non-depressed controls as well as data covering grey and white matter structure to differentiate between patients with stable vs. progressive mild cognitive impairment (MCI). There was one multimodal component differing between MDD and controls (corrected for multiple comparisons). This represented a pattern of widespread changes to white matter microstructure and associated decreases in grey matter volume overlapping with the most distinct population age effect. In further exploratory analyses we identified multimodal components related to specific MDD symptom profiles as well as inflammation. Multimodal analyses helped to disentangle widespread effects observed in conventional univariate single-modality analyses, for example by separating grey matter volume changes into multiple clinically interpretable patterns. In progressive MCI we identified a multimodal representation following established patterns of neurodegeneration. A wide range of diagnostic modelling approaches aiming at optimisation and adjustment for the specific training dataset dimensionality and diagnostic question of interest was evaluated. In the larger depression dataset the main approach applied support vector machines combined with automated feature selection in an extensive parameter optimization framework. This, however, only yielded near chance level diagnostic accuracies (up to 55.6 %). Further models using few selected components still performed far below clinical relevance. In MCI, the single best multimodal component reached an area under the curve of 0.71 in a receiver operating characteristics (ROC) analysis in order to predict progression, which also is below potential clinical relevance. In conclusion, joint multimodal modelling did not help achieve a clinically applicable diagnostic performance in mental disorders. However, multimodal modelling yielded highly interpretable representations and helped further disentangle widespread effects observed in standard univariate analyses.
Publications
- Developmental venous anomalies can interact with local functional MRI measures. Neuroradiology 2017, 59 (Suppl1):S43. Annual meeting of the European Society of Neuroradiology, Malmö, Sweden
Benedikt Sundermann, Bettina Pfleiderer, Heike Wersching, Klaus Berger, Gwenaëlle Douaud
(See online at https://doi.org/10.1007/s00234-017-1872-5)
