Medical Diagnosis and Treatment Recommendations (Doc-Bot) using Large Language Models

Abstract

This paper presents the development of an all-inclusive healthcare system designed to support medical professionals and empower patients with medical perception. The system combines conventional machine learning (ML) methods with cutting-edge models of large language (LLMs), like GPT-3 and GPT-4. It addresses various medical queries, including disease symptoms, treatment options, medication information, and preventive care advice. Additionally, it predicts diseases based on patient-reported symptoms using a symptom-to-disease mapping model trained on healthcare datasets. The disease prediction model was fine-tuned on 6,800 samples representing 135 diseases, achieving a 98% accuracy in just 12 epochs while keeping a tight eye on training loss to prevent overfitting. The model was trained with an ideal sample ratio using an NVIDIA T4 GPU to guarantee reliable performance. The system's overall performance was evaluated using metrics such as accuracy, loss monitoring, and learning rate optimization. Testing on a benchmark dataset of clinical scenarios revealed an 85% accuracy in providing correct preliminary diagnoses and a 92% relevance score for answering general medical queries. These results highlight the potential of the system to deliver accurate diagnoses and reliable recommendations, demonstrating significant potential for improving patient education and assisting medical professionals in routine diagnostic engagements while maintaining user trust and adherence to ethical standards.