Adaptive wavenumber selection framework for medicine authenticity assessment

The proliferation of counterfeit medicines presents a critical challenge to global health. ATR-FTIR spectroscopy offers a rapid and efficient means of analyzing these products, but the high dimensionality of spectral data requires effective wavenumber selection. This paper introduces a novel two-step adaptive wavenumber selection framework for classifying authentic and falsified medicines. The method initially partitions spectral data into intervals based on class distance, followed by an iterative ranking process that integrates both wavenumber relevance and redundancy. By penalizing correlated features, our approach avoids redundant information and identifies optimal wavenumber combinations. When applied to the Cialis dataset, the proposed approach achieved near-perfect accuracy (99.97 %), retaining only 2.9 wavenumbers; as for the Viagra dataset, the method achieved 98.73 % accuracy while retaining a subset of 12.4 wavenumbers. Compared to the relevance-only alternative approach, our method retained significantly fewer wavenumbers across all classifiers while consistently achieving comparable or even higher classification performance.