Generative AI in Healthcare: A Turkish Chatbot for Symptom Assesment and Tailored Recommendations

Chatbots are computer programs designed to simulate conversation with human users, typically through textual or auditory methods. In the medical domain, chatbots play a crucial role in enhancing healthcare delivery by providing immediate and personalized responses to patient queries. They serve as virtual assistants capable of offering medical advice, scheduling appointments, delivering healthcare information, and even triaging symptoms. The importance of chatbots in healthcare lies in their ability to improve access to medical expertise, especially in remote or underserved areas, reduce wait times for non-emergency consultations, and empower patients to make informed decisions about their health. In this thesis, we have developed a locally-operated chatbot designed specifically for the medical domain using large language models (LLMs). The system, named MedTRAG (Turkish Medical Text Retrieval with Constructive Learning and Knowledge Refinement), has been successfully implemented to respond to patient inquiries with relevant information and recommend appropriate doctors based on their reported complaints. By harnessing the capabilities of LLMs, we have achieved significant improvements in the efficiency and accuracy of medical consultations in an automated environment. Our work specifically addresses the challenges faced by low-resource languages in developing effective retrieval systems, with a particular focus on enhancing medical text retrieval in Turkish. To this end, we introduced a hybrid architecture combining dense retrieval and generative components through ColBERT and LLaMA 3.1 models. A specialized Turkish medical knowledge base containing 5,000 curated examples and a comprehensive triplet dataset were developed to support training of the embedding model. The system also incorporates a domain-specific knowledge representation component that transforms user queries into medically meaningful representations, further improving response precision. Experimental evaluations demonstrated substantial performance gains in Turkish medical text retrieval, with our system achieving 72.65 points in the Exam for Specialization in Medicine (TUS)—a score that would qualify for admission into 89.2\% of medical programs at Hacettepe University, one of Turkey's top medical institutions. This confirms the practical utility and reliability of our system in high-stakes academic and clinical environments. Additionally, we reduced the generation of incorrect information through semantic similarity verification techniques. Recent studies have highlighted the effectiveness of prompt engineering and fine-tuning techniques, which we incorporated to further enhance chatbot performance. Furthermore, by ensuring the chatbot operates without the need for internet connectivity, we address critical security concerns related to private data, such as vaccine studies and proprietary medical research. This local deployment ensures that confidential information remains securely within the organization, mitigating risks associated with cloud-based systems. This research has successfully bridged the gap between patients and healthcare providers, improved access to medical guidance in the Turkish language, and delivered a secure, contextually aware, and high-performing chatbot framework tailored to the needs of both users and healthcare institutions.