ATR-FTIR Coupled With Machine Learning Provides a Fast Method for Identifying and Distinguishing 55 Varieties of Fruit-Derived Medicinal Materials.

Abstract

Introduction: Fruit-derived medicinal materials (FDMM) are extensively utilized in daily life, yet the market is beset by substantial variety confusion, which undermines consumer rights and well-being. Consequently, accurate identification of these materials is essential for guaranteeing their quality, effectiveness, and safety.

Objectives: This study aimed to combine attenuated total reflection-Fourier transform infrared spectroscopy (ATR-FTIR) and machine learning (ML) techniques to differentiate and identify 55 kinds of FDMM.

Materials and methods: A total of 861 sample batches were collected, with 721 allocated for model establishment and 140 for independent validation. A PLS-DA model alongside nine machine learning algorithms-namely support vector machine (SVM), tree, K-nearest neighbor (KNN), discriminant, ensemble, support vector machine kernel (SVMK), logistic regression kernel (LRK), naive Bayes (NB), and neural network (NN)-were constructed. Considering both accuracy and computational efficiency, the optimal model was selected and evaluated in terms of its accuracy, precision, recall, and F1-score. The optimal model was further validated using 140 newly collected samples to ensure its long-term stability after several months.

Results: Among the 10 classification models, the KNN model showed exceptional classification capability, with all evaluation metric exceeding 0.98. The KNN model was validated by the new 140 samples with a prediction accuracy of 85.7%, confirming its capability in distinguishing most FDMM.

Conclusion: The application of ATR-FTIR technology combined with the robust classification capabilities of ML models enabled rapid and accurate differentiation and identification of 55 FDMM, which contributed to ensuring their quality.