A9 - Scalable Medical AI at the Edge: Stacking Strategies for Real-Time, On-Device Diagnostics

Introduction: : Deploying Large Language Models (LLMs) in clinical settings presents significant challenges, including high computational demands, limited availability of labeled data, and the need for frequent task-specific fine-tuning. In this work, we present an edge-to-edge optimized, meta-learning-enhanced LLM architecture—comprising approximately 7 billion parameters—capable of delivering high-accuracy, low-latency medical diagnoses across diverse clinical tasks with minimal supervision. In contrast to conventional approaches that utilize retrieval-augmented generation (RAG), our framework leverages a hybrid LLM ensemble fine-tuned via Model-Agnostic Meta-Learning (MAML). This design enables rapid adaptation to new healthcare tasks without the need for comprehensive retraining.

Materials and

Methods: : Our system integrates four LLMs within a unified multimodal framework and employs a hybrid stacking pipeline, where deep learning and LLM-derived embeddings are fused at the base level and further optimized at the meta level using logistic regression and XGBoost. The resulting architecture is compact and adaptable, specifically engineered for on-device inference and real-time learning in both clinical and home health environments.

Results, Conclusions, and Discussions:: We evaluated the proposed framework across three critical domains—hypertension risk assessment, cancer imaging analysis, and neurodegenerative disease monitoring—utilizing real-world, multimodal datasets curated through AI-driven pipelines. The model demonstrated robust diagnostic performance (AUC: 0.9170–0.9899) with minimal task-specific retraining required. These findings illustrate that a MAML-optimized LLM stack can provide a scalable, efficient alternative to traditional RAG pipelines for personalized, always-available medical AI. Our approach establishes a foundation for low-power, real-time diagnostic systems, facilitating the translation of research-grade LLMs into practical, deployable clinical tools.

Acknowledgements and/or References (Optional)::