A Resource-Efficient LLM-RAG Framework for Robust Clinical Decision Support

Abstract

The rapid growth of medical data and increasingly complex clinical workflows call for more capable and efficient decision support systems. Although recent language model–based approaches show promise, they often lack sufficient medical specialization and place heavy demands on computational resources. Lightweight retrieval-augmented methods offer a partial solution, yet their performance in nuanced clinical scenarios remains limited. To address these issues, we present the Adaptive Enhanced Clinical Intelligence Assistant (AECIA), a retrieval-augmented framework tailored for reliable and resource-efficient clinical decision support. AECIA integrates three core components: an adaptive fine-tuning strategy that allocates parameter-efficient updates according to layer sensitivity, a context-aware retrieval module that combines multi-granularity chunking with dynamic retrieval and re-ranking, and a prompt orchestration mechanism that adjusts prompt structure based on query characteristics and metadata cues. Built on a compact instruction-tuned model, AECIA achieves clear improvements on medical understanding and reasoning benchmarks, particularly in challenging specialties such as advanced medicine and genetics. These results demonstrate that adaptive fine-tuning, targeted retrieval, and flexible prompting can markedly strengthen clinical decision support while remaining feasible on modest hardware.