A Locally Executable AI System for Improving Preoperative Patient Communication: Multidomain Clinical Evaluation

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing
• Medical Informatics

Background:

• Patients undergoing invasive procedures often experience anxiety and have unmet information needs.
• Current large language models (LLMs) face challenges in healthcare deployment, including hallucinations, data privacy concerns, and high energy consumption, limiting equitable access.

Purpose of the Study:

• To develop and evaluate LENOHA (Low Energy, No Hallucination, Leave No One Behind Architecture), a locally executable dialog system.
• To enable safe, equitable, and sustainable preprocedural communication for patients.

Main Methods:

• Expert-curated FAQ databases were created for tooth extraction and gastroscopy.
• A sentence-transformer classifier routed clinical questions to FAQs (nongenerative) and casual conversation to a local small language model (generative).
• Evaluated sentence-transformer models against cloud LLMs (ChatGPT, Gemini Advanced) for accuracy and measured on-device energy consumption.

Main Results:

• The E5-large-instruct model achieved 98.3% accuracy, comparable to ChatGPT (GPT-4o) with 98.5% accuracy.
• The nongenerative clinical path used significantly less energy (2.23 mWh) compared to the generative small talk path (168.27 mWh), a 75-fold difference.
• Low-energy, local hardware achieved high-precision clinical support without cloud dependency.

Conclusions:

• Nongenerative clinical information retrieval is feasible on local hardware, enhancing safety and privacy.
• Decoupling clinical information from generative AI reduces energy use and improves equity in medical AI deployment.
• LENOHA offers a practical model for sustainable and accessible AI-driven patient communication in healthcare.