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

We constructed a system that explicitly separates clinically oriented queries from low-risk casual conversation using a sentence-transformer classifier. Clinical inputs are routed to a non-generative, expert vetted Frequently Asked Questions (FAQ) retrieval module that returns predefined answers verbatim, while non-clinical inputs are handled by a locally hosted small language model (SLM) restricted to brief, empathic small talk. Performance tests on a standard laptop GPU demonstrated that the routing module achieved over 98% accuracy (AUC 0.996) across two domains (tooth extraction and gastroscopy), with decision thresholds calibrated by Youden's index. Within the FAQ-based clinical pathway, all test utterances mapped to pre-authored question-answer pairs, so clinical replies remained deterministic and expert-approved rather than being generated. System profiling revealed that the non-generative FAQ pathway reduced energy consumption by approximately 75x (2.23 mWh vs. 168 mWh per request) and achieved substantially lower latency (∼0.10 s vs. ∼8.51 s) compared to the generative SLM pathway. These findings support the feasibility of deploying a privacy-preserving, energy-efficient hybrid architecture for preoperative patient communication that operates entirely offline, and they provide an implementation-focused complement to our broader research on the clinical and ethical aspects of such systems.