ABSTRACT NUMBER: ESOC2026YS5 A SAFETY-ORIENTED MULTIMODAL SMARTPHONE EXAMINATION TOOL FOR EARLY STROKE RECOGNITION: DEVELOPMENT AND INTERIM EVALUATION OF A GUIDED CLINICAL EXAM–BASED AI SYSTEM

Abstract Background and aims AI tools for stroke screening often rely on facial asymmetry alone, which may miss strokes without obvious droop or posterior-circulation deficits. We developed a brief, guided, smartphone-based neurologic exam integrating multiple signs to support rapid recognition while preserving a conservative triage approach. We designed and measured a workflow that minimizes time-to-EMS activation by stopping the exam early at the first abnormality, while extending neurologic coverage to posterior strokes through a structured exam sequence. Methods A <60-second guided workflow was created using computer vision and acoustic models. Modules include facial asymmetry detection, ocular landmark–based gaze deviation estimation, tongue/eyebrow movement symmetry, pose-based arm drift analysis, and speech clarity assessment. Feasibility testing was conducted in healthy volunteers (n=30). Interim model performance was evaluated using annotated video samples. Results All participants completed the exam successfully (median time-to-EMS-call was 14 seconds). Early multimodal performance showed AUROC values of 0.92 for facial signs, 0.88 for gaze/tongue/brow, and 0.86 for arm/speech tasks. In simulated triage scenarios, the “any abnormality → EMS” rule produced no false-negative outputs. Longer screening times occur only in cases where: (1) symptoms are subtle or non-classical, or (2) the relevant exam item is positioned later in the sequence (e.g., gaze or coordination for posterior strokes). Preliminary design of the smartphone tool is feasible, user-friendly, and offers broader neurological coverage than facial asymmetry alone. The safety-first design aligns with prehospital triage principles by avoiding probability-based reassurance. Prospective testing in suspected stroke and clinical validation is planned. Conflict of interest Hung Phan Huu: nothing to disclose; Tam Quoc Minh Tran: nothing to disclose; Nhan Le Thanh: nothing to disclose; Danh Vo: nothing to disclose; Minh Le: nothing to disclose