Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Automated acute aphasia screening using transcribed connected speech
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Lokesha Pugalenthi1, Tatiana Schnur2; 1Rice University, Houston, USA, 2McGovern Medical School, University of Texas Health Science Center, Houston, USA
Aphasia, a language disorder most often caused by a left hemisphere (LH) stroke, disrupts language at multiple levels, including retrieving single words, combining words into phrases, and producing utterances that relate coherently across a narrative (i.e., connected speech or discourse). Thus, aphasia hinders one’s ability to communicate, work, and live independently. As discourse is central to real-world communication, comprehensive aphasia evaluations typically include discourse assessment. At the acute stage of stroke (<1 week post-stroke), discourse assessments are often unfeasible because patients are medically unstable and fatigued, while clinicians are overburdened with high caseloads. Consequently, clinicians prioritize evaluating swallowing function (dysphagia) and the ability to communicate immediate needs (functional communication). Compounding these challenges, existing discourse assessments can be time-intensive and insufficiently validated for acute settings. Therefore, there is a need for a brief, discourse-based aphasia screener that can be administered and analyzed in acute care without overextending patients or clinicians. Here, to screen for the presence or absence of aphasia in acute stroke, we applied Natural Language Processing (NLP), large language models (LLMs), and machine learning (ML) classification algorithms to 1-2 minute story retellings from 65 participants with acute LH stroke (average 4 days post-stroke, 40 with aphasia). First, using manually transcribed story retellings, we applied NLP to derive 21 linguistic features previously demonstrated to be sensitive to discourse deficits in aphasia at the word- (n=10), phrase- (n=8), discourse- levels (n=3). These linguistic features were used as input to ML classification algorithms to build a linguistic classifier. The linguistic classifier achieved 74% balanced accuracy (68% sensitivity and 80% specificity; most accurate classification algorithm: logistic regression) in predicting the presence or absence of aphasia in acute LH stroke. Second, we used four LLMs (GloVe, BERT, Mistral, OpenAI) to encode the transcribed story retellings into vector representations of semantic meaning (embeddings), where semantically similar samples (e.g., “Cinderella is a girl”, “Cinderella is a woman”) have similar embeddings. Each LLM’s embedding classifier achieved balanced accuracy scores ranging from 68% to 78%. Combining the four embedding-based classifiers into an ensemble classifier outperformed the linguistic classifier, achieving 86% balanced accuracy (75% sensitivity and 96% specificity; most accurate classification algorithm: shallow neural network). Third, adding the linguistic classifier to the ensemble yielded only a small improvement to 87% balanced accuracy (82% sensitivity and 92% specificity; most accurate classification algorithm: logistic regression). This study developed a brief, automated, and clinically feasible screener of transcribed real-world speech that identifies aphasia in patients with acute stroke. It extends previous work on discourse in stroke patients by distinguishing between patients with and without aphasia, and by using an ecologically valid discourse task (story retellings) to support individual-level assessment in the acute stage. Overall, these findings represent an important step toward integrating discourse-based measures into acute clinical language assessments.
Topic Areas: Disorders: Acquired, Computational Approaches