Detecting Post-Stroke Aphasia Via Brain Responses to Speech in a Deep Learning Framework

TL;DR

This study targets the need for objective, time-efficient aphasia assessment after stroke by using neural tracking of natural speech in EEG. A deep learning framework pre-trains convolutional networks on healthy individuals to model intact acoustic, segmentation, and linguistic speech processing, then applies these models to individuals with aphasia and controls. An SVM classifier, fed with 11 CNN-derived neural-tracking measures plus age, achieves 85.4% accuracy (AUC 86.2%) for individual-level aphasia detection, with robust performance after only about 9 minutes of EEG data. The results show widespread reductions in neural tracking in IWA across all speech representations, highlighting the potential for an automated, generalizable, and time-efficient screening tool in clinical settings, while underscoring the need for broader validation and extension to more severe cases.

Abstract

Aphasia, a language disorder primarily caused by a stroke, is traditionally diagnosed using behavioral language tests. However, these tests are time-consuming, require manual interpretation by trained clinicians, suffer from low ecological validity, and diagnosis can be biased by comorbid motor and cognitive problems present in aphasia. In this study, we introduce an automated screening tool for speech processing impairments in aphasia that relies on time-locked brain responses to speech, known as neural tracking, within a deep learning framework. We modeled electroencephalography (EEG) responses to acoustic, segmentation, and linguistic speech representations of a story using convolutional neural networks trained on a large sample of healthy participants, serving as a model for intact neural tracking of speech. Subsequently, we evaluated our models on an independent sample comprising 26 individuals with aphasia (IWA) and 22 healthy controls. Our results reveal decreased tracking of all speech representations in IWA. Utilizing a support vector machine classifier with neural tracking measures as input, we demonstrate high accuracy in aphasia detection at the individual level (85.42\%) in a time-efficient manner (requiring 9 minutes of EEG data). Given its high robustness, time efficiency, and generalizability to unseen data, our approach holds significant promise for clinical applications.

Figures (3)

• Figure 1: A. Match-mismatch task. Finding which of the matched (yellow) or mismatched (black) segments is synchronized with the EEG segment. B. Model architectures. All models contain an EEG segment, a matched, and a mismatched segment of speech. They are fed to a CNN block, compared with cosine similarity, and passed through a last dense layer to select the matched segment. The single-feature architecture contains only the top stream, indicated with the red rectangle, while the dual-feature architecture includes the entire scheme.
• Figure 2: Match-mismatch accuracy values grouped for all features.A. Accuracy values for all envelope-based features. B. Accuracy values for segmentation features (phoneme and word onsets) and linguistic features, grouped per level (phoneme/word level). Group comparisons were performed using Wilcoxon rank-sum tests. ***: $p<.001$, **: $p<.01.$
• Figure 3: SVM performance and feature importance ranking.A. ROC curve, plotting the false positive rate as a function of the true positive rate. B. SHAP analysis investigating the impact of each feature on the SVM's predictions. Features are ranked according to their impact. C. SVM performance as a function of EEG recording length. The 9-minute mark, achieving the same accuracy as the entire recording, is indicated in gray.