Traditional Machine Learning Models Versus Small Language Models For Classification of Orthoses, Prostheses, and Special Materials In Invoices

This paper investigates the use of traditional machine learning models versus Small Language Models (SLMs) for classifying Orthoses, Prostheses, and Special Materials (OPMEs) in invoices, a critical problem for healthcare auditing. Building on the OPME dictionary used by the OPMinEr system, a controlled experiment is conducted comparing a Linear SVM classifier with TF–IDF representations against a fine-tuned BERTimbau model, in both a binary task (OPME vs non-OPME) and a multiclass task (orthosis, prosthesis, special material, others). The results show that the SVM achieves accuracy and macro F1 scores around 0.98–0.99, slightly outperforming BERTimbau in both tasks, while requiring orders of magnitude less training and inference time. These findings suggest that, in the specific context of healthcare invoices, well-tuned traditional models remain the most effective and cost-efficient choice for large-scale deployment, whereas SLMs are most promising as complementary tools in research settings and for potential transfer to related classification tasks.