CAM-CDSS: Context-Augmented and Med-Adapted Lightweight Language Models for Enhanced Clinical Decision Support

Abstract

The immense potential of large language models for clinical decision support systems (CDSS) faces challenges like insufficient domain specificity, high resource consumption, and critical data privacy concerns. To overcome these limitations, we propose Context-Augmented & Med-Adapted CDSS (CAM-CDSS), a novel lightweight, efficient, and trustworthy system. CAM-CDSS refines the Retrieval-Augmented Generation paradigm by integrating an enhanced base model with multi-stage domain-adaptive fine-tuning for broad medical and task-specific alignment. It incorporates a semantic-enhanced hybrid retrieval mechanism, leveraging entity recognition, relation extraction, and structured metadata for precise context acquisition, alongside intelligent context formulation, dynamic prompt engineering, and a continuous knowledge update. Experimental results demonstrate that CAM-CDSS consistently outperforms baseline models across diverse medical benchmarks, successfully mitigating performance regressions observed in challenging domains. Our analysis highlights CAM-CDSS's efficiency for consumer-grade deployment, enhanced retrieval, and superior human evaluation scores in correctness, relevance, and clinical acceptability. CAM-CDSS offers potential for accurate, robust, and trustworthy decision support.

Ähnliche Arbeiten