Dear Editor,
In a recent article, Boehm et al. combine functional magnetic resonance imaging (fMRI) with classification-based multivariate pattern analysis [1] to investigate changes in the cognitive representations of visually presented food stimuli in anorexia nervosa (AN) patients who were acutely underweight (acAN) and in weight-recovered AN patients (recAN) [2]. The authors report that a machine learning algorithm can discriminate representations of food stimuli better than representations of neutral stimuli—in certain brain regions—within the acAN sample but not within an age-matched sample of healthy controls (HCacAN). Moreover, this discriminability of food-vs-neutral representations in the recAN sample does not statistically differ from that of an age-matched sample of healthy controls (HCrecAN). The authors take these findings to suggest that cortical representations of food are altered in acAN compared to recAN, which may be indicative of altered attentional mechanisms in the presence of food in AN patients.
Such a pattern of results may also be useful as a prognostic marker, based on the authors’ regression analysis after a one-year assessment following treatment, thereby offering fMRI and cognitive neuroscience methods an additional avenue to inform the clinical domain. As such, this study is a great example of investigating cognitive processes with functional neuroimaging in a patient population and should act as a stepping stone upon which future studies can build. To this end, there are three aspects of the article that merit further discussion: caveats in the statistical inference, complementing the classification analysis with representational similarity analysis [3], and the manner in which attentional mechanisms may underlie such results.
The authors’ main finding from the multivariate analysis involves a difference in classification performance for food-vs-neutral representations when contrasting acAN patients with HCacAN, and that this difference in classification performance diminishes when contrasting recAN with HCrecAN. However, the contrast that warrants the most meaningful interpretation of the results (given the scope of the study) is the interaction of these two contrasts (i.e., [acAN > HCacAN] > [recAN > HCrecAN]). Reporting evidence for a difference between one set of groups (e.g., p1 < 0.05) and a lack of evidence for a difference between the other set of groups (e.g., p2 > 0.05) merely compares their effect sizes but does not directly test for the difference between the group differences [4]. Instead, demonstrating evidence for the interaction, and crucially that the interaction is driven by the contrast [acAN > recAN], would provide the most compelling evidence for the authors’ interpretation of the results (Fig. 1a, b). Additionally, given that the control samples also differed in their ages, finding that the interaction is not driven by the contrast [HCacAN > HCrecAN] would help to rule out the between-groups age-confound mentioned in the limitations section, as the authors would demonstrate that age alone is insufficient to explain potential differences between acAN and recAN. Note that this criticism does not imply that the authors’ interpretation is necessarily incorrect, but rather that the interpretation (i.e., that recAN differs from acAN) is not directly warranted from the statistical tests performed; in the best case, the patient samples differ from one another, while the control samples do not (i.e., [acAN ≠ recAN] ∩ [HCacAN = HCrecAN]), while in the worst case, the opposite pattern is observed (i.e., [acAN = recAN] ∩ [HCacAN ≠ HCrecAN]).
Regardless, exploring differences in classification performance to infer altered information processing for prognostic purposes is a shrewd application of machine learning [5]. However, given the different goals of decoding-based and encoding-based analyses [6], one can complement the classification analyses (Fig. 1c, d) with, for example, representational similarity analysis [7], thereby gleaning insight regarding how the representations are changing [8].
This strategy would be of particular interest, given the authors’ supposition that altered attentional mechanisms towards food underlie the classifier’s differential performance across groups. Previous work has shown that attention alters representational spaces to increase the categoricity of the attended feature [9]. As such, one could investigate whether neural representations of food in acAN patients are, for example, unusually dispersed (or compact) with respect to those of healthy controls or recAN (Fig. 1e, f) and, with additional encoding analyses, whether these representations tend to distribute along different dimensions underlying the food representational space. Such approaches could help to unravel how attentional mechanisms may (pathologically) affect cognitive processes related to food in acAN patients and determine whether any individual-level alterations in the representational space have additional prognostic value [10].
This correspondence aims to highlight the clever manner in which the authors combined machine learning with fMRI to investigate cognitive changes in patients with AN while simultaneously drawing attention to a few caveats/strategies in the analyses and interpretations that researchers and reviewers should take into account when designing and assessing functional neuroimaging experiments.
References
Norman KA, Polyn SM, Detre GJ, Haxby JV. Beyond mind-reading: multi-voxel pattern analysis of fMRI data. Trends Cogn Sci. 2006;10:424–30.
Boehm I, Mohr H, King JA, Steding J, Geisler D, Wronski M-L, et al. Aberrant neural representation of food stimuli in women with acute anorexia nervosa predicts treatment outcome and is improved in weight restored individuals. Transl Psychiatry. 2021;11:532.
Kriegeskorte N, Mur M, Bandettini P. Representational similarity analysis—connecting the branches of systems neuroscience. Front Syst Neurosci. 2008;2:1–28.
Nieuwenhuis S, Forstmann BU, Wagenmakers E-J. Erroneous analyses of interactions in neuroscience: a problem of significance. Nat Neurosci. 2011;14:1105–7.
Janssen RJ, Mourão-Miranda J, Schnack HG. Making individual prognoses in psychiatry using neuroimaging and machine learning. Biol Psychiatry Cogn Neurosci Neuroimaging. 2018;3:798–808.
Kriegeskorte N, Douglas PK. Interpreting encoding and decoding models. Curr Opin Neurobiol. 2019;55:167–79.
Weaverdyck ME, Lieberman MD, Parkinson C. Tools of the Trade Multivoxel pattern analysis in fMRI: a practical introduction for social and affective neuroscientists. Soc Cogn Affect Neurosci. 2020;15:487–509.
Levine SM, Pfaller M, Reichenberger J, Shiban Y, Mühlberger A, Rupprecht R, et al. Relating experimentally-induced fear to pre-existing phobic fear in the human brain. Soc Cogn Affect Neurosci. 2018;13:164–72.
Nastase SA, Connolly AC, Oosterhof NN, Halchenko YO, Guntupalli JS, Visconti di Oleggio Castello M, et al. Attention selectively reshapes the geometry of distributed semantic representation. Cereb Cortex. 2017;27:4277–91.
Levine SM, Schwarzbach JV. Individualizing representational similarity analysis. Front Psychiatry. 2021;12:1727.
Acknowledgements
The author would like to thank Philipp Seidel for comments on a previous version of this manuscript.
Author information
Authors and Affiliations
Contributions
SML is the sole author of this work.
Corresponding author
Ethics declarations
Competing interests
The author declares no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Levine, S.M. Probing patterns for prognostic potential. Transl Psychiatry 12, 167 (2022). https://doi.org/10.1038/s41398-022-01931-z
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41398-022-01931-z
This article is cited by
-
The challenge to disentangle age and recovery effects in research focusing on adolescent anorexia nervosa
Translational Psychiatry (2022)