Algorithm and hyperparameter optimizations for hetero-device classification by near-infrared spectra of falsified and substandard amoxicillin capsules

Abstract

In this work, we optimized classification algorithms and the hyperparameters for screening falsified and substandard amoxicillin capsules. The distribution of low-quality medical products is a serious problem, especially in low- and middle-income countries. Near-infrared (NIR) spectroscopy has been proposed as the first choice for a screening device. However, preparation of the reference library for the classification training is a highly difficult process. We herein propose a hetero-device classification between training and test devices. In this proposal, Fourier-transform NIR spectrometer and portable wavelength dispersive NIR spectrometer were used as training and test devices, respectively. As the classifier candidates, we examined 13 algorithms and selected 8. We then optimized the hyperparameters for these classifiers by the grid search and cross validation methods. In the final analysis, few classifiers were found to give acceptable prediction results by the hetero-device classification. When using these methods, it is crucial to examine the results by the classification probability, due to the trade-off between sensitivity and specificity. Finally, we suggest that k-nearest neighbors, extra trees, and gradient boosting classifiers are the optimal algorithms with high classification probability for the substandard and falsified amoxicillin capsules.