Automated Medical Device Classification

Abstract

Manual categorization of medical devices according to regulatory guidelines such as CDSCO (Central Drugs Standard Control Organization) is generally time-consuming, variable, and susceptible to human error. This article describes an automated system of classification based on Microsoft's MiniLM-L12-H384-uncased model, fine-tuned using a dataset of device descriptions, classes, and respective departments. The model is enhanced further by using TF-IDF vectorization and cosine similarity for the determination of the nearest matching device. Experimental evidence shows better classification accuracy and robustness, enabling this technique for large-scale deployment in a regulatory context.