Análise exploratória e experimental de aplicações de inteligência artificial para classificação de descrições incongruentes em compras na área de saúde pública

Abstract

Context: The Orthoses, Prostheses and Special Materials (OPME) sector in the health area presents a wide variety of products and technologies, involving both multinational companies as local. Despite technological advances, many services and information systems, especially in the public sector, still use unstructured descriptions in natural language of products, services or events, making classification and analysis difficult. However, for efficient audits, it is necessary to automatically classify and total invoices issued for purchasing products. In this way, the lack of standardization in the nomenclature in commercialization of OPMEs, not only makes it difficult to compare products, either for standardizing prices or standardization of use, but it also opens up space for possible acts of corruption. Objective: To mitigate the problem of ineffective standardization and coding, develop and evaluate the effectiveness and efficiency of an OPME classifier, in the context of electronic invoice descriptions, from the perspective of auditors, health professionals and data scientists. Method: Initially, Systematic Mapping (SM) was carried out as a way to identify and characterize the Artificial intelligence approaches and techniques for automatically classifying incongruent textual descriptions on invoices. Subsequently, an artificial intelligence-based tool, OPMinEr, was implemented to classify OPME invoices. A controlled experiment was then carried out to evaluate the mapped Artificial Intelligence (AI) algorithms. Results: The search strategy used in the systematic mapping selected 225 articles, which passed the inclusion and exclusion criteria. Among the approaches found to solve the problem of incongruent textual descriptions, supervised machine learning was prominent in 60% of the works. In the controlled experiment, considering statistical significance, the Linear Support Vector algorithm achieved an accuracy of 99% and stood out among the others. In terms of efficiency, the algorithm Naïve Bayes Multinomial stood out, having the fastest average training time, with 4.375 seconds. Conclusion: The results demonstrate that it is possible to automatically identify and classify OPMEs in invoices, allowing for a more precise analysis of indicators such as anomalously high prices and quantities of OPMEs purchased per inhabitant. These analyses are conducted by the Audit of the Unified Health System (AudSUS), Ministry of Health - Brazil, to identify potential irregularities and contribute to transparency and efficiency in healthcare resource management.