Background: Large vessel occlusion (LVO) stroke comprises 30-37% of all strokes and contributes disproportionately to stroke morbidity (61.6%) and mortality (95.6%). Aphasia is experienced in c.30% of LVO stroke and has a high sensitivity for LVO stroke prediction (91%). Emerging studies have applied artificial intelligence (AI) to predict LVO stroke and address challenges in accurate pre-hospital triage. Early detection of LVO stroke is critical to ensure access to time-critical interventions.
Aim: To evaluate the feasibility of using AI to differentiate people with and without stroke from speech.
Methods: This prospective observational study included consenting adults with a recent diagnosis of acute LVO stroke and controls without any neurological conditions. Ethics approval was obtained (South Western Sydney Local Health District Human Research Ethics Committee, 2022/ETH00483). Data included demographic and clinical information and a short video taken on an iPhone while administering National Institutes of Health Stroke Scale (NIHSS) “Best Language” and “Dysarthria” questions. Supervised and unsupervised machine learning was used to detect whether audio (speech) data could differentiate stroke.
Results: Participants included 10 with LVO stroke (median age=71.6 (IQR=52-78)) and 10 controls. AI models could differentiate stroke from non-stroke: GPT-4 achieved a sensitivity and specificity of 0.6 and 1.0 (Task 1), and the wav2vec2 models achieved an AUC of 0.89 and 0.97 (Tasks 2-4). Several words and phrases demonstrated greater differentiation (“cactus,” “mama,” “huckleberry”, “near the table in the dining room”).
**Conclusion: **Using multiple AI algorithms to predict LVO stroke based on language and acoustic features is feasible. "