This scientific commentary refers to ‘High-dimensional therapeutic inference in the focally damaged human brain’, by Xu et al. (doi:).
In Europe, more than 2 million individuals each year will have their brain integrity and function challenged by stroke. Despite the progress achieved through the use of thrombolysis with alteplase in acute stroke patients (Wahlgren et al., 2007), many have persistent deficits, affecting their personality, degrading their quality of life and preventing their return to work. Patients need to know, in a timely manner, to what extent their symptoms will resolve. This knowledge reduces the burden and stress associated with stroke, and allows patients to make appropriate arrangements with their family, their health insurance provider and their employer. The mechanics of recovery also provide a framework in which to study the organization of the human brain. Despite the large body of statistically significant lesion-symptom studies in the literature, there has been relatively little translation of these studies into clinical practice, mainly because of their weak predictive power. In effect, the methods currently employed in clinical neuroanatomical studies of stroke are unfit for clinical practice. Research is currently failing practitioners and patients alike. Radical changes in clinical-neuroanatomical correlations are required in order to make significant progress in the field of stroke research. In this issue of Brain, Xu and co-workers provide novel insights into the therapeutic inferences that can be derived from a focally damaged human brain (Xu et al., 2017).
The authors used a large dataset of stroke patients (n = 1172), with no subselection of the population other than exclusion of very large lesions involving an entire hemisphere to avoid potential bias. As there is no ground truth regarding the impact of any therapeutic intervention, Xu et al. instead simulated interventions and their potential effects on a specific outcome, namely gaze deviation after stroke. The authors used two types of simulated interventions: a ‘lesion non-altering intervention’, which changes the brain’s response to the lesion but not the lesion per se (e.g. neurorehabilitation), and a ‘lesion-altering intervention’, which reduces the size of the lesion (e.g. thrombolysis). In both cases, after iteratively randomizing the patients to ‘intervention’ and ‘control’ cohorts, Xu et al. tested whether statistical models could detect different effect sizes of treatment. A high-dimensional multivariate approach exhibited a significantly higher predictive power than that of the classical low-dimensional approach for the non-altering intervention. This effect was amplified for the lesion-altering intervention. The work of Xu et al. thus demonstrates the biases inherent in the classic lesion-symptom approach, and offers a statistical solution to help investigators draw clinically meaningful inferences in future.
Thanks to data-sharing initiatives, any investigator can replicate the methods of Xu et al. or apply them to a different question. Such initiatives have provided the scientific community with a wealth of data for analysis. For instance, the Enigma Consortium stroke recovery initiative (http://enigma.ini.usc.edu/ongoing/enigma-stroke-recovery/) recently made available 304 T1-weighted MRIs with manually segmented lesions and metadata (Liew et al., 2017). Many authors now recognize the importance of analysing large numbers of unselected stroke patients in order to increase sensitivity, rather than drawing conclusions derived from carefully selected groups and exceptional clinical cases (Corbetta et al., 2015).
There are several approaches available for modelling the interactions between anatomical regions (Fig. 1), including statistical modelling (Xu et al., 2017), and use of a priori anatomical knowledge derived from diffusion-weighted imaging tractography (disconnectome approaches, http://toolkit.bcblab.com). Statistical modelling (i.e. a high-dimensional multivariate approach) captures interactions between damaged areas across patients and can help avoid bias (Xu et al., 2017). Other authors have used pre-established atlases of white matter connections to estimate the association between damaged regions (Thiebaut de Schotten et al., 2015). This approach has identified biomarkers with high sensitivity and specificity for chronic visuospatial neglect (Thiebaut de Schotten et al., 2014). Finally, functional MRI at rest (resting state functional MRI) can be used to measure the spontaneous association between brain areas (i.e. functional connectivity) (Boes et al., 2015; Foulon et al., 2017). Recent results demonstrate that higher order cognitive symptoms, such as impairments in memory and attention, are better explained by changes in functional connectivity than by lesion location, whereas lesion location better explains primary disorders (e.g. motor and visual disorders) (Siegel et al., 2016). This finding suggests that the highest level of cognitive function emerges from the interaction between primary and lower level areas through a mechanism of integration, which is not localized to a single brain area. The dynamics of functional connectivity closely reflect the state of recovery within the brain, which in turn makes resting state functional MRI an excellent method for studying the mechanisms associated with brain recovery. However, functional connectivity is not ideal for predicting future symptoms, as it fluctuates during the recovery process.
Three models capture the association between brain regions in clinical-neuroanatomical correlations. Analyses were reduced to three patients (Patients p1, p2 and p3) and two voxels (v1, v2) for simplification purposes and performed with BCBtoolkit (http://toolkit.bcblab.com). (A) Mass multivariate approach whereby all voxels of the brain are considered variables of interest. The interaction between lesioned voxels is integrated into the statistical model used to predict patient outcomes. (B) Disconnectome maps, whereby the probability of disconnection in each voxel is considered a variable of interest. Unlike the mass-multivariate approach, voxels are tested independently in the statistical model used to predict patient outcomes. (C) Functional connectivity maps, whereby the correlation between the average time course of the BOLD signal in the lesion and each of the other voxels in the brain is considered a variable of interest. The voxels are tested independently in the statistical model used to predict patient outcomes.
There are significant challenges ahead in improving approaches based on lesion-symptom statistics, diffusion-weighted imaging tractography, and resting state functional connectivity. A framework that takes advantage of all three methods could provide a more comprehensive understanding of the biology of recovery after stroke. For instance, tractography can reveal the long-range effects of a brain lesion, whereas functional connectivity measures the dynamic changes associated with brain recovery through plasticity. As proposed by Xu et al., high-dimensional multivariate statistics can control the relationship between various dimensions under study and the distribution of the lesion in space to avoid any bias and optimize the ability of the model to detect interactions. This framework could be used to revisit classic syndromes and provide biomarkers with large effect sizes, which would be applicable in the clinic. Another challenge will be to quantify the percentage of shared variance between the model and individual measurements. Most of these methods assume that all brains are equivalent or identical (Thiebaut de Schotten and Shallice, 2017). Preliminary evidence suggests that different brain connectivity phenotypes are associated with different profiles of recovery. For instance, Forkel et al. showed that the size of the arcuate fasciculus in the right hemisphere predicts some of the improvements in speech impairment after a stroke in the left hemisphere (Forkel et al., 2014). Again, reporting the effect size of this interindividual variability is crucial in order to estimate its importance in everyday clinical practice.
In conclusion, associationist principles (Geschwind, 1965; Catani and ffytche, 2005) indicate that brain lesions should not be considered random focal impairments, but rather anatomically predetermined imbalances in a vast system of interconnected areas. Previous attempts to build on this framework were limited by the available methodology and by a lack of access to large cohorts of patients. Newer ideas involving optimized statistics, anatomical models and functional measures have already demonstrated their validity and superiority over classical lesion overlapping approaches in predicting symptoms and the likely outcomes of clinical interventions. We now must reach out and engage with other scientific teams, and promote the sharing of technologies and datasets, to build up a common framework that will be beneficial to patients and to society as a whole.
Funding
We thank Lauren Sakuma for discussion and the “Agence Nationale de la Recherche” [grant numbers ANR-13-JSV4-0001-01 and ANR-10-IAIHU-06] for their generous support.
