The research, published in the journal Frontiers in the Internet of Things, comes from the Federal University of Sergipe and the Laboratory of Technological Innovation in Health at the Federal University of Rio Grande do Norte. The team set out to solve a problem that has long frustrated auditors at Brazil’s Ministry of Health: the OPME sector, which spans everything from stents and catheters to orthopaedic prostheses and cannulas, is dominated by a bewildering variety of products from both multinational giants and small local firms. Because many public information systems still record purchases as unstructured natural language descriptions rather than standardized codes, comparing products, standardizing prices and detecting anomalies becomes a manual, error-prone chore. Worse, that lack of standardization opens space for corruption.

The stakes are enormous. The researchers cite estimates that corruption costs Brazil between 1.38 and 2.30 percent of its gross domestic product, which translates to roughly 100 to 160 billion reais based on 2019 figures. The OPME market itself gained national notoriety in January 2015, when the press exposed an alleged fraudulent scheme involving manufacturers, distributors, hospitals, doctors and lawyers, which became known as the prostheses mafia. The scheme reportedly involved selling devices at inflated prices, paying irregular commissions and outright embezzlement. The problem is not unique to Brazil: in 2017, the European Court of Auditors reported that misclassification and misdescription of goods are among the most common forms of payment evasion.

At the heart of the new tool, called OPMinEr, is a classic but powerful text-mining technique known as TF-IDF, short for Term Frequency-Inverse Document Frequency. The method works by splitting each invoice description into tokens, the individual words separated by spaces. Term frequency counts how often a word appears in a given document, while inverse document frequency measures how representative that word is across the entire corpus of documents. The final TF-IDF weight multiplies these two values, so words that appear in nearly every document receive a weight of zero, since the logarithm of one is zero. In other words, common words carry little discriminatory power, while rare, distinctive terms like specific device names stand out. This vector representation allows machine learning algorithms to treat each invoice description as a point in a high-dimensional numerical space.

Building the knowledge model required an ambitious data curation effort. The researchers assembled a data dictionary from multiple sources: product records from Brazil’s National Health Surveillance Agency (ANVISA), the federal government’s official materials register (CATMAT), descriptions entered manually by auditors, and invoices from the state of Rio Grande do Norte that were labeled by hand. A Python script automated the downloading of CATMAT files, though the team had to contend with the instability of the federal government’s website, which periodically stopped responding and forced the script to retry downloads. To semi-automatically identify OPME products within the ANVISA and CATMAT databases, the team used cosine similarity, a natural language processing technique that measures how closely two text descriptions resemble each other. Descriptions with a similarity score of 0.6 or less were initially discarded, and the remainder underwent manual review before being added to the dictionary. The team also created a standardization and conflation dictionary that acts as a synonym tool, automatically correcting truncated words like parafus into the complete form parafuso, meaning screw.

OPMinEr performs three sequential classifications. First, it predicts whether a product description corresponds to an OPME at all; if not, it is labeled as Other. Next, it assigns the product to one of 83 different classes, such as Stent, Catheter, Orthopaedic Prosthesis or Cannula. Finally, it predicts the specific Unified Health System (SUS) procedure in which the product may have been used, drawing on 319 possible procedures ranging from Detachable Balloon Placement to Tibio-Tarsal Intramedullary Stem. Each classification type required its own trained model, since different algorithms may excel at different tasks.

To evaluate the tool scientifically, the team conducted a controlled in vitro experiment using 465,726 electronic invoices from January 2020 to May 2022, obtained from AudSUS, the SUS National Audit Department, as well as from LAIS and the Federal Public Ministry. Because invoices for products that are not OPMEs vastly outnumber OPME invoices, the researchers selected a sample of 4,718 records that preserved the approximate proportion of OPME notes. Seven algorithms were compared: Multinomial, Gaussian and Bernoulli Naive Bayes, Linear Support Vector, Decision Tree, Random Forest and Gradient Boosting. The models were trained and evaluated using 10-fold cross-validation, meaning the data was divided into ten parts and the model was trained and tested ten times, with metrics averaged across iterations. The team used PostgreSQL, Python, Django, MLflow, Power BI and scikit-learn on a computer with an Intel Core i7 processor and 32 gigabytes of RAM.

The results were striking. For the Type classification, the Linear Support Vector algorithm achieved an accuracy of 99.8 percent, along with matching sensitivity, precision and F1-score of 99.8 percent, while training in just over seven seconds and classifying in 0.273 seconds. Random Forest followed closely at 99.6 percent accuracy, though it took over 221 seconds to train. Gradient Boosting, despite good accuracy, proved painfully slow, taking more than four days to train the Procedure classification data, because the algorithm builds a new decision tree in each iteration to correct the mistakes of previous trees by assigning higher weights to misclassified examples. For the Class classification, Linear Support Vector again led at 99.6 percent accuracy, and for the Procedure classification it reached 98.6 percent. Multinomial Naive Bayes was the most efficient in average training time across all three classification types, while Decision Tree was the fastest at average classification time. Bernoulli Naive Bayes collapsed on the more complex tasks, dropping to 73.5 percent accuracy for Class and a dismal 37.1 percent for Procedure, with precision of just 13.8 percent in the latter case.

The team did not rely on raw numbers alone. They applied five statistical tests: Shapiro-Wilk to check normality of the data distribution, Kruskal-Wallis and Post-Hoc Dunn for non-normal data, and ANOVA with Tukey HSD for normal distributions. With a significance level of 0.05, the tests confirmed that differences between algorithms were statistically significant. After applying the Bonferroni correction, Linear Support Vector proved significantly different from all three Naive Bayes variants in the Type classification. Interestingly, although Linear Support Vector had the highest accuracy, the statistical tests showed that Linear Support Vector and Random Forest, which approach the problem quite differently, one seeking the hyperplane that maximizes the margin between classes, the other averaging votes from many decision trees trained on different data subsets, were statistically similar in performance.

The practical payoff comes in the form of audit trails and intelligent reports for AudSUS auditors. Once invoices are classified, the system can flag anomalously high prices, unusual quantities of OPMEs purchased per inhabitant, and suspicious relationships between requesting and authorizing professionals. In one example, the system detected that out of 17,861 authorizations, 2,223, or 12.4 percent, were made by one doctor to another, with 100 percent of the first doctor’s authorizations going to the second, totaling over 4.1 million reais. Another dashboard identifies cases where the quantity of a material purchased by a health unit significantly exceeds expected demand, suggesting cold invoices or stolen stock. The researchers note that the classifier is a sine qua non condition for executing OPME audit trails, since without accurate classification, no downstream analysis is possible. Looking ahead, the team plans to explore deep learning approaches, including a hybrid metamodel incorporating BERT, a transformer-based language model that considers both preceding and following context of words, potentially offering an even more complete understanding of invoice descriptions. For now, the Linear Support Vector classifier, chosen for its superior accuracy and its ability to maximize margins between classes for strong generalization, stands as a working weapon in the fight for transparency in health resource management.

Subject of Research: A smart classifier of orthoses, prostheses and special materials (OPMEs) in invoices

Article Title: A smart classifier of orthoses, prostheses and special materials (OPMEs) in invoices

Article References: Gomes, W., Colaço Júnior, M., Alves, L. P. C. F., Fontes, R., Silva, R., Nunes, B., Silva, C., & Valentim, R. (2025). A smart classifier of orthoses, prostheses and special materials (OPMEs) in invoices. Frontiers in the Internet of Things, 4, Article 1436757. https://doi.org/10.3389/friot.2025.1436757

Image Credits: AI Generated

DOI: 10.3389/friot.2025.1436757

Keywords: artificial intelligence, machine learning, healthcare fraud, text classification, public auditing, Brazil, smart, classifier, orthoses, prostheses, special, materials