Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks

Cluster-based delineations of aphasia profiles using EEG measures of acoustic and linguistic speech encoding

There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.

Poster A54 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port

Jill Kries¹, Pieter De Clercq¹, Ramtin Mehraram¹, Robin Lemmens^2,3,4, Tom Francart¹, Maaike Vandermosten¹; ¹Experimental Oto-Rhino-Laryngology, Department of Neuroscience, Leuven Brain Institute, KU Leuven, Leuven, Belgium, ²Experimental Neurology, Department of Neurosciences, KU Leuven, Leuven, Belgium, ³Laboratory of Neurobiology, VIB-KU Leuven Center for Brain and Disease Research, Leuven, Belgium, ⁴University Hospitals Leuven, Department of Neurology, Leuven, Belgium

Unsupervised machine learning methods can help us develop the diagnostic tools of tomorrow by analyzing clinical data patterns and grouping patients based on their clinical profiles (Alashwal et al., 2019). This data-driven approach may complement already existing diagnostic categories that have previously been established based on case studies. Such is the case for aphasia, a language disorder for which categorization of subtypes goes back to the days of Broca and Wernicke. Although there is a shift towards using more data-driven methods to find patterns of aphasia subtypes (Wilson & Hula, 2019), the classical aphasia typology is still used in the clinic to date. Nonetheless, accumulating evidence shows that a neurobiologically informed diagnosis - as opposed to a diagnosis based on behavioral language tests - may represent the large heterogeneity in aphasia phenotypes more accurately (Pasley & Knight, 2013; Tremblay & Dick, 2016). Here we clustered 41 individuals with aphasia (IWA) in the chronic phase after stroke based on their outcomes on a natural speech listening paradigm while EEG data was recorded. Specifically, the relationship between speech features and the EEG signal was studied by means of encoding models. We combined several speech features in 4 encoding models to achieve outcome measures that represent (1) acoustic processing, (2) speech segmentation-related processing, (3) linguistic processing at phoneme level and (4) linguistic processing at word level (see Kries et al., biorxiv, https://doi.org/10.1101/2023.03.01.530707 for details). Before clustering, we scaled the data using the RobustScaler from Scikit-learn and conducted principal component analysis to reduce dimensions from 4 speech encoding models to 2 principal components (n=2 was pre-defined). The cluster analysis was performed using Scikit-learn’s k-means clustering. To choose the optimal number of clusters, we looked at silhouette scores (i.e., reflecting how close a score is to its own cluster and how far it is from another cluster) and at within-cluster sum of squared errors (via the elbow method). Together, the conclusions from both methods showed that a division of IWA into 3 clusters and into 7 clusters would be optimal. Hence, we conducted a feature analysis for both options. When dividing IWA into 3 clusters, we observed, via visual inspection, that the 3 clusters are distinguishable in amplitude of speech encoding across dimensions, but also that there are differences between dimensions. For example, IWA in ‘cluster0’ seem to have relatively higher encoding of phoneme-level linguistics (surprisal and entropy), while IWA in ‘cluster2’ seem to have relatively lower encoding of phoneme-level linguistics and segmentation cues (phoneme and word onsets), but relatively higher encoding of acoustic cues (envelope and envelope onsets). Similar, but more complex patterns within and across speech encoding dimensions were observed when IWA were delineated into 7 clusters. This preliminary analysis, which will be further optimized in the coming months, shows that machine learning algorithms may be a valuable asset to delineate data-driven profiles of aphasia. However, a larger dataset in the future would certainly yield more robust outcomes. This study presents a first step towards a neurobiologically informed, data-driven diagnostic tool for aphasia.

Topic Areas: Disorders: Acquired, Computational Approaches