JMIR Publications

ABSTRACT

Background:

Large language models (LLMs) offer promising capabilities for medical communication but face significant barriers in clinical deployment, including data privacy concerns, the risk of hallucinations, and high energy consumption associated with cloud-based infrastructure. Existing solutions often prioritize performance over sustainability and local implementability.

Objective:

The aim of this study was to develop and evaluate "LENOHA" (Low Energy, No Hallucination, Leave No One Behind Architecture), a locally executable AI system designed to ensure data privacy, prevent hallucinations, and minimize energy consumption while maintaining high conversational accuracy in preoperative settings.

Methods:

We constructed a hybrid system running on a local consumer GPU (NVIDIA RTX 3080). The architecture routes clinical queries to a non-generative FAQ matcher (Sentence Transformer) and casual conversation to a local small language model (SLM). We evaluated the system using expert-supervised synthetic datasets across two domains: tooth extraction and gastroscopy. Performance metrics included classification accuracy, response latency, and power consumption (mWh/request) compared to a generative-only approach.

Results:

The system achieved a classification accuracy of 0.983, which was statistically indistinguishable from GPT-4o. The non-generative clinical path consumed approximately 1.0 mWh per request, whereas the generative path required approximately 168.4 mWh, representing a ~170-fold difference in energy efficiency. The system effectively eliminated hallucinations for clinical queries by design.

Conclusions:

LENOHA demonstrates that a locally executable, hybrid AI architecture can achieve high performance and safety while significantly reducing energy consumption and eliminating data privacy risks. This study provides a sustainable model for deploying medical AI in resource-constrained environments. Clinical Trial: N/A