A Dual-Modality Clinical Decision Support Framework: From Foundation Models to Deployable Conversational AI

Progress in deep learning is still a hurdle for the translation of clinical NLP models into deployable healthcare applications. Although the hybrid LLM-GNN models have demonstrated their ability in diagnostic tasks, there is a need to assess their capabilities in the conversational domain, computational efficiency, and validation for their practical application. This study builds upon a dual-modality architecture of Clinical-LLM-GNN that has been validated in previous studies, with two goals. First, we build a medical conversational AI system, in which we classify the intent (84.38% accuracy, 82.66% Macro F1 across 8 intents), recognize named entities (5 entity types), handle multi-turn dialogue (5-turn context), and generate responses based on the knowledge graph (8 diseases, 5 drugs), with a response time of 1.35ms. Second, we perform extensive validation by ablating studies (8 configurations), statistical testing (6 comparisons), and complexity analysis. The hybrid model BERT+GGNN+Attention excels in terms of accuracy (99.87%) with significant gains (p<0.05) over baselines, while retaining practical computational footprint (113.1M parameters, 47.3ms inference). The framework supports SDG 3 and SDG 9 by providing easy to use and scalable clinical decision support, which is a step towards deployment of conversational AI that improves clinical workflow and patient information access using foundation models.