Emerging Properties Exploration with UMAP Statistics: EMUSE
Poster No:
1904
Submission Type:
Abstract Submission
Authors:
Chris Foulon1, Marcela Ovando-Tellez2, Lia Talozzi3, Anna Matsulevits4, Fanny Munsch5, Igor Sibon5, Thomas Tourdias6, Michel Thiebaut de Schotten7
Institutions:
1Groupe d'Imagerie Neurofonctionnelle (GIN), Institut des Maladies Neurodegeneratives-UMR 5293, CNRS, Bordeaux, France, 2Institut des maladies neurodégénératives (IMN), Bordeaux, France, 3Stanford Medical School, Stanford, CA, 4University Bordeaux, Institut des Maladies Neurodégénératives CNRS UMR 5293 Université de Bordeaux, Bordeaux, Gironde, 5Université de Bordeaux, Bordeaux, France, 6University Bordeaux, Bordeaux, Gironde, 7Groupe d’Imagerie Neurofonctionnelle, Institut des Maladies Neurodégénératives- UMR 5293, CNRS, CEA, Bordeaux, France
First Author:
Chris Foulon
Groupe d'Imagerie Neurofonctionnelle (GIN), Institut des Maladies Neurodegeneratives-UMR 5293, CNRS
Bordeaux, France
Groupe d'Imagerie Neurofonctionnelle (GIN), Institut des Maladies Neurodegeneratives-UMR 5293, CNRS
Bordeaux, France
Co-Author(s):
Anna Matsulevits
University Bordeaux, Institut des Maladies Neurodégénératives CNRS UMR 5293 Université de Bordeaux
Bordeaux, Gironde
University Bordeaux, Institut des Maladies Neurodégénératives CNRS UMR 5293 Université de Bordeaux
Bordeaux, Gironde
Michel Thiebaut de Schotten
Groupe d’Imagerie Neurofonctionnelle, Institut des Maladies Neurodégénératives- UMR 5293, CNRS, CEA
Bordeaux, France
Groupe d’Imagerie Neurofonctionnelle, Institut des Maladies Neurodégénératives- UMR 5293, CNRS, CEA
Bordeaux, France
Introduction:
With the improvement of in-vivo human brain imaging techniques and analyses, it is becoming increasingly clear that cognitive functions are not localised but emerge from the complex interaction between brain areas[1]. Therefore, new tools are required to capture these emergent properties.
Uniform Manifold Approximation and Projection (UMAP)[2], a dimensionality reduction technique, is increasingly popular in displaying data structure. Specifically, UMAP creates a morphospace representing different patterns in the input data relative to each other with a proportional distance representing their similarity or dissimilarity-ideal to identify how combining different factors can lead to a unique result. Accordingly, Tallozi et al. proposed a statistical method to associate the disconnection patterns in the UMAP morphospace with psychological deficits[3] that demonstrated high predictive performance and robustness[4].
Here, we propose a new open-source Python tool-EMUSE-that extends the same framework to a more general setting, allowing anyone to build their morphospace and explore its organisation statistically. In a proof-of-concept example, we show that we can associate stroke lesion profiles with cognitive deficits.
Uniform Manifold Approximation and Projection (UMAP)[2], a dimensionality reduction technique, is increasingly popular in displaying data structure. Specifically, UMAP creates a morphospace representing different patterns in the input data relative to each other with a proportional distance representing their similarity or dissimilarity-ideal to identify how combining different factors can lead to a unique result. Accordingly, Tallozi et al. proposed a statistical method to associate the disconnection patterns in the UMAP morphospace with psychological deficits[3] that demonstrated high predictive performance and robustness[4].
Here, we propose a new open-source Python tool-EMUSE-that extends the same framework to a more general setting, allowing anyone to build their morphospace and explore its organisation statistically. In a proof-of-concept example, we show that we can associate stroke lesion profiles with cognitive deficits.
Methods:
We used the delineated lesion masks in the standard MNI152 space of 187 stroke patients[5]. We train the UMAP model that represents the profile of the lesions, placing similar lesions close together and different lesions far apart in a continuous and unbounded 2D space (Fig 1a). To run statistics on this space, we create a 2D pixelated space with customisable parameters-number of pixels and smoothing-where we can expand the coordinates to represent proximity and explore how this proximity relates to the examined measures (Fig 1b). Then, we ran a pixel-wise Pearson correlation (corrected for multiple comparisons and a p-value threshold of 0.05) and showed the parts of the UMAP space associated with significant variations in the studied psychological test. Finally, we can retrieve which lesions relate to the behavioural changes.
We analysed which lesion profiles are associated with a deficit in language and motor tasks in the UMAP morphospace based on the 187 lesion masks. The language task is part of the Montreal Cognitive Assessment (MoCA)[6], and the motor score is part of the Fugl-Meyer test[7]. We chose these two tests to show our method can capture both primary and high-level cognitive impairments that would emerge from complex lesion patterns.
We analysed which lesion profiles are associated with a deficit in language and motor tasks in the UMAP morphospace based on the 187 lesion masks. The language task is part of the Montreal Cognitive Assessment (MoCA)[6], and the motor score is part of the Fugl-Meyer test[7]. We chose these two tests to show our method can capture both primary and high-level cognitive impairments that would emerge from complex lesion patterns.
Results:
We were able to capture the lesion profiles of both motor and language deficits in our dataset (Fig 2). After correction for multiple comparisons, lesion profiles correlated with a motor deficit between 0.32 and 0.39 and a language deficit between 0.318 and 0.324. The comparison of the overlap of the lesions contributing to the correlated lesion profiles in the UMAP morphospace with the NeuroSynth[8] decoding tool shows that the lesions are consistent with the literature. The lesion profiles associated with a motor deficit were correlated with multiple subcortical structures involved in motor tasks-putamen, basal ganglia, caudate nucleus, striatum-and the insula. The lesions associated with a language deficit were correlated with the superior and middle temporal gyrus and the auditory cortex, as well as the concepts of listening, sounds and speech.
Conclusions:
Our findings demonstrate that EMUSE is not only able to identify the interaction between complex lesion patterns accurately but also proficiently establish statistical correlations between these interactions and cognitive deficits. This tool has been meticulously crafted for accessibility and user-friendliness, thereby empowering researchers to seamlessly generate and investigate various UMAP morphospaces tailored to their unique datasets. Furthermore, the versatility of EMUSE allows for adaptation to a wide array of data inputs and uses of various statistical techniques, making it a novel asset in neuroscientific research.
Modeling and Analysis Methods:
Methods Development 1
Multivariate Approaches 2
Other Methods
Neuroinformatics and Data Sharing:
Informatics Other
Keywords:
Computational Neuroscience
Informatics
Machine Learning
Modeling
Multivariate
1|2Indicates the priority used for review
Provide references using author date format
[2] McInnes, L., Healy, J., & Melville, J. (2018). 'Umap: Uniform manifold approximation and projection for dimension reduction.' arXiv preprint arXiv:1802.03426.
[3] Talozzi, L., Forkel, S. J., Pacella, V., Nozais, V., Allart, E., Piscicelli, C., ... & Thiebaut de Schotten, M. (2023). 'Latent disconnectome prediction of long-term cognitive-behavioural symptoms in stroke.' Brain, 146(5), 1963-1978.
[4] Hope, T. M., Neville, D., Talozzi, L., Foulon, C., Forkel, S. J., de Schotten, M. T., & Price, C. J. (2023). 'Testing the Disconnectome Symptom Discoverer model on out-of-sample post-stroke language outcomes.' Brain: a journal of neurology, awad352.
[5] Munsch, F., Sagnier, S., Asselineau, J., Bigourdan, A., Guttmann, C. R., Debruxelles, S., ... & Tourdias, T. (2016). 'Stroke location is an independent predictor of cognitive outcome.' Stroke, 47(1), 66-73.
[6] Nasreddine, Z. S., Phillips, N. A., Bédirian, V., Charbonneau, S., Whitehead, V., Collin, I., ... & Chertkow, H. (2005). 'The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment.' Journal of the American Geriatrics Society, 53(4), 695-699.
[7] Fugel-Meyer, A. R., Jaasko, L., Leyman, I., Ollson, S., & Steglind, S. (1975). 'The post-stroke hemiplegic patient1, a method for evaluation of physical perpormance.' Scand. J. Rahabil. Med, 7, 13-31.
[8] Yarkoni, T., Poldrack, R. A., Nichols, T. E., Van Essen, D. C., & Wager, T. D. (2011). 'Large-scale automated synthesis of human functional neuroimaging data.' Nature methods, 8(8), 665-670.