Image-based identification and DEA-based optimization modeling of antibiotic packaging using unsupervised learning techniques

Abstract

Background Ensuring medication safety requires accurate identification of antibiotic packaging, especially within pharmacy automation and dispensing systems. Advanced imaging and machine learning offer novel avenues for physical package recognition. Objective To investigate visual and textual features of antibiotic packages and evaluate their relationship with identification outcomes using unsupervised learning and efficiency-based analysis. Methods Thirty-six antibiotic formulations from Thailand (2016–2021) were analyzed using binary imaging, entropy metrics, packaging area ratio (PAR), and optical character recognition (OCR). K-means clustering was applied to segment package groups, and data envelopment analysis (DEA) was used to assess relative efficiency without assuming predefined functional relationships between inputs and outputs. Results Nine distinct image clusters were identified. Packages with mid-range entropy (7.1–7.5) and PAR (1.2–1.45) were associated with higher identification consistency. OCR text confidence influenced identification outcomes. DEA identified clusters with relatively efficient input–output configurations. Conclusion Integrating image-derived metrics and OCR-based features supports automated antibiotic package identification. This framework provides a structured approach for evaluating packaging characteristics in pharmacy workflows.