I have included the summary from Neuroskeptic and the beginning of the full article - follow the links below to see the original article.
By Neuroskeptic | July 1, 2014
A provocative and important paper just out claims to have identified a pervasive flaw in many attempts to map the function of the human brain.
University College London (UCL) neuroscientists Yee-Haur Mah and colleagues say that in the light of their findings, “current inferences about human brain function and deficits based on lesion mapping must be re-evaluated.”
Lesion mapping is a fundamental tool of modern neuroscience. By observing the particular symptoms (deficits) people develop after suffering damage (lesions) to particular parts of the brain, we can work out what functions those various parts perform. If someone loses their hippocampus, say, and gets amnesia, you might infer that the function of the hippocampus is related to memory – as indeed it is.
However, there’s a problem with this approach, Mah et al say. Conventional lesion mapping treats each point in the brain (voxel) individually, as a possible correlate of a given deficit. This is called a mass univariate approach.
The problem is that the shape and location of brain lesions is not random – some areas are more likely to be affected than others, and the extent of the lesions varies in different places.
What this means is that the presence of damage in a certain voxel may be correlated with damage in other voxels. So damage in a voxel might be correlated with a given deficit, even though it has no role in causing the deficit, just because it tends to be damaged alongside another voxel that really is involved.
Mah et al call this problem ‘parasitic association’. In a large sample of diffusion-weighted MRI scans from 581 stroke patients, the authors show that the co-occurrence of damage across voxels leads to systematic, large biases in mass univariate deficit mapping.
The biases follow a complex geometry throughout the brain: as this lovely (but scary) image shows -
This shows the direction and magnitude of the error that would afflict a standard attempt to localize a hypothetical deficit that was truly associated with points throughout the brain. In some areas, the bias ‘points’ forward, so the deficit would be wrongly mapping as being further forward than it really was. In other places, it points in other direction.
The mean size of the expected error is 1.5 cm, but with a high degree of variability, so it is much worse in some areas.
Worse yet, Mah et al say that in cases where the same deficit can be caused by damage to two, non-adjacent areas, univariate lesion mapping might fail to pinpoint either of them. Instead, it could wrongly implicate a nearby, unrelated area.
The authors conclude that the only way to avoid this problem is by using multivariate statistics to explicitly model voxel interrelationships, e.g. a machine learning approach. This will require large datasets, but they caution that merely having a big sample, without multivariate statistics, would achieve nothing. They conclude on a somewhat downbeat note:
It is outside the scope if this study to determine the optimal multivariate approach: our focus here is on the evidence of the misleading picture the mass-univariate approach has created, and the need to review it wholesale. Taken together, our work demonstrates a way forward to place the study of focal brain lesions on a robust theoretical footing.
I’m not sure the outlook is quite so bleak. It’s a very nice paper, however, Mah et al’s dataset is purely based on stroke patients. Yet there are many other sources of brain lesions that are used for lesion mapping: tumours, infections, and head injuries to name a few.Full Citation:
These kinds of lesions probably throw up parasitic associations as well, however, they might be very different from the kind seen in strokes. This is because strokes, unlike other kinds of lesions, are always centered on blood vessels.
Mah et al note that the bias map they found is clustered around the major cerebral arteries and veins, but the obvious conclusion to draw from this is that it’s only applicable to strokes.
Whether other kinds of lesions produce substantial biases remains to be established. Until we know that, we shouldn’t be rushing to redraw any maps just yet.
Mah, Y., Husain, M., Rees, G., & Nachev, P. (2014, Jun 28). Human brain lesion-deficit inference remapped. Brain; DOI: 10.1093/brain/awu164
Included here is the abstract and the introduction - follow the link in the title to download the PDF for yourself.
Yee-Haur Mah, Masud Husain, Geraint Rees, and Parashkev Nachev
1. Institute of Neurology, UCL, London, WC1N 3BG, UKSummary
2. Department of Clinical Neurology, University of Oxford, Oxford OX3 9DU, UK
3. Institute of Cognitive Neuroscience, UCL, London WC1N 3AR, UK
4. Wellcome Trust Centre for Neuroimaging, UCL, London WC1N 3BG, UK
Our knowledge of the anatomical organization of the human brain in health and disease draws heavily on the study of patients with focal brain lesions. Historically the first method of mapping brain function, it is still potentially the most powerful, establishing the necessity of any putative neural substrate for a given function or deficit. Great inferential power, however, carries a crucial vulnerability: without stronger alternatives any consistent error cannot be easily detected. A hitherto unexamined source of such error is the structure of the high-dimensional distribution of patterns of focal damage, especially in ischaemic injury—the commonest aetiology in lesion-deficit studies—where the anatomy is naturally shaped by the architecture of the vascular tree. This distribution is so complex that analysis of lesion data sets of conventional size cannot illuminate its structure, leaving us in the dark about the presence or absence of such error. To examine this crucial question we assembled the largest known set of focal brain lesions (n = 581), derived from unselected patients with acute ischaemic injury (mean age = 62.3 years, standard deviation = 17.8, male:female ratio = 0.547), visualized with diffusion-weighted magnetic resonance imaging, and processed with validated automated lesion segmentation routines. High-dimensional analysis of this data revealed a hidden bias within the multivariate patterns of damage that will consistently distort lesion-deficit maps, displacing inferred critical regions from their true locations, in a manner opaque to replication. Quantifying the size of this mislocalization demonstrates that past lesion-deficit relationships estimated with conventional inferential methodology are likely to be significantly displaced, by a magnitude dependent on the unknown underlying lesion-deficit relationship itself. Past studies therefore cannot be retrospectively corrected, except by new knowledge that would render them redundant. Positively, we show that novel machine learning techniques employing high-dimensional inference can nonetheless accurately converge on the true locus. We conclude that current inferences about human brain function and deficits based on lesion mapping must be re-evaluated with methodology that adequately captures the high-dimensional structure of lesion data.
The study of patients with focal brain damage first revealed that the human brain has a functionally specialized architecture (Broca, 1861; Wernicke, 1874). Over the past century and a half such studies have been critical to identifying the distinctive neural substrates of language (Broca, 1861; Wernicke, 1874), memory (Scoville and Milner, 1957), emotion (Adolphs et al., 1995; Calder et al., 2000), perception (Goodale and Milner, 1992), decision-making (Bechara et al., 1994), attention (Egly et al., 1994; Mort et al., 2003), and intelligence (Gla¨ scher et al., 2009), casting light on the anatomical basis of deficits resulting from dysfunction of the brain. Though functional imaging has revolutionized the field of brain function mapping in the last 20 years, the necessity of a brain region for a putative function—arguably the strongest test—can only be established by showing a deficit when the function of the region is disrupted. Inactivating brain areas experimentally cannot easily be done in humans; the special cases of transcranial magnetic and direct current stimulation, though potentially powerful, are restricted temporally to days and anatomically to accessible regions of cortex.
The only comprehensive means of establishing functional necessity thus remains the study of patients with naturally occurring focal brain lesions (Rorden and Karnath, 2004). Though single patients may sometimes be suggestive, robust, population-level inferences about lesion-deficit relationships require aggregation of data from many patients (Karnath et al., 2004). Analogously to functional brain imaging, a statistical test comparing groups of patients with and without a deficit is iteratively applied point-bypoint to brain lesion images parcellated into many volume units (voxels) (Bates et al., 2003; Karnath et al., 2004). Voxels that cross the significance threshold are then taken to identify the functionally critical brain areas whose damage leads to the deficit.
Crucially, this ‘mass-univariate’ approach assumes that the resultant structure-deficit localization is not distorted by co-incidental damage of other, non-critical loci in each patient: in other words, that damage to each voxel is independent of damage to any other. This cannot be assumed in the human brain. Collaterally damaged but functionally irrelevant voxels might be associated with voxels critical for a deficit through an idiosyncrasy of the pathological process—the distribution of the vascular tree, for example—while having no relation to the function of interest. Such associations would lead to a distortion of the inferred anatomical locus.
Importantly, these ‘parasitic’ voxel-voxel associations can be detected only by examining the multivariate pattern of damage across the entire brain, and across the entire group. Studying large numbers of patients with the standard approach simply exacerbates the problem, because such consistent error will also consistently displace inferred critical brain regions from their true locations. Equally, replicating a study with the same number of patients will replicate the error too: observing the same result across different research groups and epochs offers no reassurance. Instead, the pattern of damage must be captured by a high dimensional multivariate distribution that describes how the presence or absence of damage at every voxel within each brain image is related to damage to all other voxels. The presence of ‘parasitic’ voxel-voxel associations would then manifest as a
hidden bias within the multivariate distribution, a complex correlation between individual patterns of damage apparent only in a high-dimensional space and opaque to inspection with simple univariate tools.
To illustrate the problem, consider the 2D synthetic example in Fig. 1, where damage to any part of area ‘A’ alone may disrupt a putative function of interest, but ‘B’ plays no role in this function of interest. If the lesions used to map the functional dependence on A follow a stereotyped pattern where damage to any part of A is systematically associated with collateral damage to the non-critical area B, both areas might appear to be significantly associated even if B is irrelevant to the function of interest. Crucially, if the pattern of the lesions within each patient is such (for reasons to do with factors unconnected to function) that the spatial variability of damage to B is less than to A, B will not only be erroneously determined to be critical but will have a higher significance value for such an association than A. The apparent locus of a lesion function deficit will therefore be displaced from A (the true locus) to B. Thus a hidden bias in the pattern of damage—hidden because it is apparent only when examining the pattern as a whole, in a multivariate way—distorts the spatial inference.
Whether or not such a hidden bias exists has not been previously investigated. Here we analyse the largest reported series of focal brain lesions (n = 581) to show that it does exist, and that it compels a revision of previous lesion-deficit relationships within a wholly different inferential framework.