Summarizing Dental Records Using Large Language Models

Abstract

The increasing volume of medical data, combined with limited time for reviewing patient histories, is often cognitively demanding for healthcare professionals. This challenge is especially evident in dentistry, where conditions such as developing periodontitis often require long-term monitoring across multiple visits. Providing an automated summary of patient history could reduce this workload, minimize the risk of overlooking critical information, and allow clinicians to focus more on patient care. Large language models (LLMs) offer a promising solution for generating such summaries automatically. This thesis investigates how LLMs can be applied to summarize dental record data effectively, without compromising clinical accuracy and safety.

To address this challenge, this thesis adopts the Design Science Research methodology. The goal is to develop a research artifact for Entteri Oy, a Finnish health technology company specializing in oral healthcare practice management software. The resulting artifact consists of two components: the microservice-based prototype and a domain-specific prompt. The microservice demonstrates a modular approach to implementing new AI functionality separately from existing software. The prompt, on the other hand, directs the LLMs to produce summaries that are both clinically appropriate and reflect awareness of relevant regulations.

To ensure safe testing, a synthetic dataset was created in collaboration with a dental professional, simulating real-world periodontal records. The effectiveness and suitability of the prompt were evaluated based on the AI-generated summaries, using manual analysis and an expert evaluation questionnaire completed by practicing dentists. Particular emphasis was placed on the terminological accuracy and clinical relevance of one summary selected for detailed analysis.

The results show that microservice architecture offers a modular and scalable way to implement AI-powered summarization functionality. The designed prompt successfully guided the model to produce clear and accurate summaries, while avoiding hallucinations and treatment recommendations. The expert evaluation questionnaire also provided valuable insights for future development. Based on these findings, general-purpose LLMs, guided by a carefully crafted prompt, can support summarization tasks in dental care.