Schedule of Events | Symposia

Automated connected speech classification of language impairment in acute stroke

Poster Session E - Monday, March 31, 2025, 2:30 – 4:30 pm EDT, Back Bay Ballroom/Republic Ballroom

Lokesha Pugalenthi¹ (lp60@rice.edu), Junyi Jessy Li², Tatiana Schnur³; ¹Rice University, ²University of Texas at Austin, ³University of Texas, Houston

Using connected speech to assess language impairment quantitatively is rare due to time constraints associated with elicitation, transcription and scoring. We predicted language impairment during acute left hemisphere (LH) stroke using connected speech to discriminate between controls, patients with language impairment, and without. We analysed connected speech narratives of 74 patients with LH acute stroke (50 with language impairment; tested within an average 3.9 days post-stroke) and 85 controls (age-, education- matched to language impairment group, p’s > 0.1; AphasiaBank, MacWhinney et al., 2011). We derived 22 lexical-semantic and syntactic linguistic features and individual narrative embeddings (OpenAI text-embedding-3-large model; 2024) as input to classification algorithms to compute balanced accuracy discrimination success. Nested cross-validation determined best classification algorithm and internal hyperparameters. Leave-one-out cross-validation generalized results across participants. Permutation analysis computed each feature’s contribution to classification performance. We achieved excellent differentiation (99% balanced accuracy, Logistic Regression) between controls and acute stroke patients with language impairment. Features affecting classification performance > 5% include mean sentence length, median utterance length, number of narrative words, degree of phrase elaboration. We achieved promising discrimination (73% balanced accuracy, Support Vector Classifier) between LH with and without language impairment. Features affecting classification performance > 5% include words per minute, mean word length, proportion closed-class to narrative words produced, number narrative words, and morphological complexity of tensed main clause verbs. Our results constitute an important step toward an automated assessment of language impairment from connected speech to facilitate referrals for speech-language therapy in the setting of acute stroke.

Topic Area: LANGUAGE: Other