System Architecture of a Local Hybrid AI for Secure Patient Communication: Implementation on Consumer Hardware.

Area of Science:

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

Background:

• Patient communication systems often struggle to balance clinical accuracy with conversational ease.
• Generative AI models raise concerns regarding data privacy, energy consumption, and response reliability in medical contexts.

Purpose of the Study:

• To design and evaluate a hybrid AI architecture for preoperative patient communication.
• To differentiate clinical queries from casual conversation for optimized response pathways.
• To enhance privacy, energy efficiency, and response determinism in patient-facing AI systems.

Main Methods:

• A sentence-transformer classifier was used to distinguish clinical from non-clinical user inputs.
• Clinical queries were routed to a non-generative FAQ retrieval module for expert-vetted answers.
• Non-clinical queries were handled by a locally hosted small language model (SLM) for brief interactions.

Main Results:

• The classifier achieved over 98% accuracy (AUC 0.996) in distinguishing clinical domains.
• The FAQ pathway provided deterministic, expert-approved responses, avoiding generative AI risks.
• The non-generative FAQ pathway demonstrated a 75x reduction in energy consumption and significantly lower latency compared to the SLM pathway.

Conclusions:

• A hybrid AI architecture is feasible for privacy-preserving, energy-efficient, offline preoperative patient communication.
• This system offers a practical implementation complementing research on clinical and ethical AI deployment.
• The approach ensures reliable and secure patient interactions by segregating query types.