Enhanced estimations of post‐stroke aphasia severity using stacked multimodal predictions

Abstract

The severity of post-stroke aphasia and the potential for recovery are highly variable and difficult to predict. Evidence suggests that optimal estimation of aphasia severity requires the integration of multiple neuroimaging modalities and the adoption of new methods that can detect multivariate brain-behavior relationships. We created and tested a multimodal framework that relies on three information sources (lesion maps, structural connectivity, and functional connectivity) to create an array of unimodal predictions which are then fed into a final model that creates “stacked multimodal predictions” (STAMP). Cross-validated predictions of four aphasia scores (picture naming, sentence repetition, sentence comprehension, and overall aphasia severity) were obtained from 53 left hemispheric chronic stroke patients (age: 57.1±12.3yrs, post-stroke interval: 20 months, 25 female). Results showed accurate predictions for all four aphasia scores (correlation true vs. predicted: r=0.79–0.88). The accuracy was slightly smaller but yet significant (r=0.66) in a full split cross-validation with each patient considered as new. Critically, multimodal predictions produced more accurate results that any single modality alone. Topological maps of the brain regions involved in the prediction were recovered and compared with traditional voxel-based lesion-to-symptom maps, revealing high spatial congruency. These results suggest that neuroimaging modalities carry complementary information potentially useful for the prediction of aphasia scores. More broadly, this study shows that the translation of neuroimaging findings into clinically useful tools calls for a shift in perspective from unimodal to multimodal neuroimaging, from univariate to multivariate methods, from linear to non-linear models, and, conceptually, from inferential to predictive brain mapping.

Structural disconnections explain brain network dysfunction after stroke

“…SDC measures were defined for each patient based on the intersection of their lesion with a structural connectome atlas constructed with data from healthy individuals (see Methods). Similar atlas-based approaches have been used by other recent lesion studies (Carter et al, 2017;Foulon et al, 2018a;Griffis et al, 2017bGriffis et al, , 2017aHope et al, 2018;Kuceyeski et al, 2013Kuceyeski et al, , 2014Kuceyeski et al, , 2015Kuceyeski et al, , 2016bPustina et al, 2017a;Ramsey et al, 2017), and analogous strategies are often employed to study SC-FC relationships in animal models (Adachi et al, 2011;Grandjean et al, 2017;Grayson et al, 2016;Shen et al, 2015b). While these approaches assume similar approximations of individual structural connectomes by the atlas and cannot account for interindividual variability in the properties of un-damaged fiber pathways (Forkel and Catani, 2018;Forkel et al, 2014), they also offer some protection against potential biases arising from interindividual differences in diffusion MRI data quality, lesion effects on data processing/reconstruction, etc.…”

Combining heterogeneous data sources for neuroimaging based diagnosis: re-weighting and selecting what is important

“…In this sense, we consider RF exploiting the I & C as segmentation of the sources of information in order to highlight the differences compared to the other presented methodologies. We implement two methods, namely Gray [20] and Pustina [21], where the RF algorithms are the key in order to find the 430 best representation of the single source of information. These methods are not kernel-based methods, and are composed by a pipeline of different algorithms.…”

“…Another possible way to combine different sources of information is using RF-based methods [20,21]. The framework used in these studies consists of several steps, where the RF methods are fundamental in order to obtain the final model as a combination of the different sources.…”

“…As another example, the algorithm in [21] consists of a sequential exploitation of graph theory, recursive feature elimination (RFE) and RF. Graph theory is used to derived a set of features that are added to them the raw data.…”

“…Secondly, we tackle the task investigated in [23] where the goal is to classify de-140 pressed patients vs healthy controls by integrating fMRI data with additional clinical information. We compare our approach with SVM [7] as the baseline approach, as well as a state-of-the-art MKL approach (SimpleMKL [4]), two fea-ture selection approaches: recursive feature elimination (RFE) [24] and t-test [25], and RF-based methods ( [20,21]).…”