Disease Prediction from Symptom Data with Context-Aware Drug Recommendation Using RAG-Based Language Models

Abstract

The timely and precise prediction of diseases based on patient-reported symptom data is a critical challenge in clinical decision support systems, particularly in scenarios requiring early intervention before laboratory results are available. This paper proposes a multi-model drug recommendation system that combines supervised machine learning with Generative Artificial Intelligence to predict illness and deliver context-aware, grounded medical recommendations. The proposed framework employs a Gradient Boosting classifier trained on structured historical patient data with symptom predictors including fever, fatigue, headache, nausea, cough, joint pain, abdominal pain, weight loss, and breathlessness. The model performs multi-class classification to predict one of forty-one disease categories spanning infectious, metabolic, neurological, and chronic conditions, achieving 100% accuracy on the evaluated test split. To translate predictive outcomes into actionable clinical guidance, a Generative AI module based on Retrieval-Augmented Generation (RAG) is integrated into the pipeline. Disease-specific medical knowledge is retrieved from curated PDF documents covering medications, disease descriptions, dietary instructions, precautions, and exercise recommendations. A LLaMA-based large language model conditions its response generation on the retrieved content, minimising hallucination and enhancing factual consistency. The modular design of the system ensures that predictive inference and recommendation generation remain independently optimisable while functioning as a unified pipeline. The experimental evaluation demonstrates the effectiveness of both the classification and generation components, confirming the viability of combining machine learning classifiers and retrieval-grounded generative models for end-to-end medical decision support in research-oriented healthcare applications.

Ähnliche Arbeiten