Abstract

International classifications are a crucial tool for organising healthcare data, helping us to understand populations both locally and globally. However, with the continuous advancement of medical knowledge, these classifications are becoming increasingly complex. For instance, the International Classification of Diseases, 11th edition (ICD-11), comprises 17,000 categories of diseases. In contrast, the International Classification of Primary Care, 3rd edition (ICPC-3), is more concise, with 1,350 codes. Despite this, healthcare professionals still face the challenge of correctly attributing codes to various clinical scenarios, making tools necessary to help them find the right codes efficiently and accurately. This research evaluates three different search engines for the ICPC-2 code, developed by the author to tackle this issue. They were created using the Python programming language and its open source libraries, and vary in their information retrieval algorithms: BM25 ranking algorithm, vector search powered by OpenAI embeddings, and a combination of both methods. They were all built with the same Portuguese thesaurus for ICPC-2 codes, which maps concepts to their appropriate codes. The search engines’ performance was evaluated and compared using their search history. The first version became available online in August 2021, only in Portuguese. All entries were reviewed by the author, assigned an expected ICPC-2 code using the thesaurus as a guide. The entries were then submitted to each information retrieval algorithm to determine if the correct code was retrieved as the first result or among the top five results. Kappa statistics was used to calculate agreement rate with respect to the reference. The goal of this presentation is to present the main results and limitations of this research, discuss the growing complexity of medical coding, explore how it can be addressed using search engines and information retrieval algorithms, and examine potential future approaches to medical coding.