Embodied Digital Therapists with LLM Personalization for Aphasia Rehabilitation: Characterizing Human-AI Collaboration Boundaries

Aphasia following stroke affects millions globally, yet rehabilitation remains severely limited by speech therapist shortages. Existing digital systems rely on static video demonstrations, single-modality assessment, and rule-based feedback, failing to address authentic clinical needs. Through formative investigation with five therapists, three patients, and their caregivers, we identified concrete clinical challenges: therapists spending 30-40% of time on repetitive demonstrations, existing tools providing only speech scores without articulatory evaluation, and patients struggling with complex interfaces and monotonous content. To address these challenges, we developed an integrated rehabilitation system combining an embodied digital therapist for Action Observation Therapy, tri-dimensional assessment coordinating speech quality, lip movement accuracy, and semantic understanding, and large language model-driven personalization for content generation and adaptive training. We conducted a proof-of-concept evaluation across two in-situ hospital training sessions with six patients, six caregivers, and three therapists. Results demonstrated substantial efficiency gains, with therapists spending 69-78% less time per patient. Patient acceptance improved 42.6% across sessions, and low digital literacy patients showed steepest gains (+69.0%). However, human intervention remained necessary for 24-30% of session time to provide emotional support. These findings empirically characterize human-AI collaboration boundaries in clinical rehabilitation, revealing both automation's potential to enhance efficiency and bridge digital divides, and the persistent necessity of human therapeutic presence—providing evidence for responsible deployment of AI-assisted healthcare systems.