Developing a Local Based LLM for Stroke (Aphasia) Rehabilitation
Approximately one-third of stroke patients suffer from Aphasia or communication disorders that result in the difficulty of speaking, writing, and reading. This project aims to develop a locally hosted AI conversation engine for GazeTalk — an augmentative and alternative communication (AAC) app designed for aphasia rehabilitation. Current GazeTalk prototype relies on cloud-based large language models (LLMs) such as GPT to power therapeutic conversations with patients. While effective, this dependency creates significant risk in terms of compliance and cost. Most importantly, a specific and local type of LLM can be more effective in aiding aphasia rehabilitation purposes.
The goal of this project is to evaluate, optimize, and integrate an open source and lighter weight LLM into GazeTalk's existing infrastructure. More specifically, the project should:
- Review existing literature and best practice on local LLM deployment in clinical and privacy-sensitive contexts (e.g. Meditron EPFL)
- Benchmark candidate open-weight models (e.g., Mistral, Llama 3, Gemma) against GazeTalk's therapeutic conversation criteria
- Design an evaluation rubric based on aphasia-specific dialogue quality (response warmth, circumlocution handling, turn naturalness)
- Fine-tune or prompt-optimize the best candidate model for therapeutic use
- Integrate the model into GazeTalk's existing Node.js server architecture via Ollama
- Conduct a comparative evaluation against the current cloud-based baseline
This project builds upon the existing GazeTalk prototype, which already runs a local Ollama pipeline and captures rich session metrics (WPM, KSPC, TTR, CIU). The student will inherit a working codebase and focus on advancing its clinical readiness. The project will be supervised by postdoctoral researcher Ekky Tammarar Alfian and co-supervised by Prof. John Paulin Hansen. This project can be formulated into a Master thesis/Bachelor thesis/student project. The project is expected to start in the Fall semester 2026 (negotiable).
Thesis type: Software implementation with benchmarking study.
Technical skills: LLM fine-tuning and prompt engineering; backend development (Node.js); evaluation methodology.
Research skills: Natural Language Processing; Human-Computer Interaction; Clinical Informatics.
Educational background: Computer Science, Artificial Intelligence, Software Engineering
ECTS credits: MSc thesis / BSc thesis / Special course (5 or 10 ECTS)
Contact: < email supprimé pour raison de sécurité >
By working on the project, you will be part of a team aiming to improve lives for stroke patients!