A Generative AI–Driven Clinical Decision Support Framework Using Large Language Models

Abstract

Early disease detection aids in the correct diagnosis and treatment of illnesses. A Clinical Decision Support System (CDS) helps identify illnesses and choose the best course of therapy. This paper presents a Generative AI-powered Clinical Decision Support Architecture based on Large Language Models (LLMs) to predict diseases and support diagnoses. The suggested architecture incorporates both structured clinical information and high-level preprocessing, feature selection, and class-balancing algorithms to increase the predictive accuracy. Experiments were conducted on 400 patient records from the UCI Chronic Kidney Disease (CKD) dataset. GPT-4o was used to learn more complex clinical patterns and aid in diagnostic decision-making. The recommended framework performed well, as evidenced by the accuracy of 99.17, sensitivity of 99.98, specificity of 98.70, F1-score of 98.85, Matthews Correlation Coefficient (MCC) of 98.21, and AUROC of 0.996. These findings are far more effective than conventional ML models and currently available LLM-based clinical methods. The high sensitivity yields a low rate of false negatives, which is essential in the early detection of disease, whereas the high specificity lowers the wrong diagnosis of healthy patients. Altogether, the suggested generative AI-based solution is powerful, consistent, and effective in clinical contexts, which underscores the potential of large language models (LLM) in medical decision support systems of the next generation.