Multimodal data fusion and multi-objective optimized CatBoost models for origin traceability and active compound prediction of Salvia miltiorrhiza

The quality of traditional Chinese medicines is strongly influenced by geographical origin; however, existing methods for origin traceability and quality evaluation still struggle to balance analytical efficiency, operational convenience, and information completeness. In this study, 120 samples of Salvia miltiorrhiza (SM) were collected from five representative production regions in China. Computer vision (CV) was used to extract appearance features, including color, texture, and contour characteristics, while high-performance liquid chromatography (HPLC) was employed to determine the contents of active components. By integrating machine learning with multivariate statistical analysis, origin classification models were developed using single-modality data as well as data-level, feature-level, and decision-level fusion strategies, followed by prediction of chemical component contents based on CV features. Results from 50 repeated stratified random splits showed that the Categorical Boosting (CatBoost) model under the decision-level fusion strategy achieved an accuracy of 99.17% ± 1.29% for origin identification, demonstrating strong classification performance and stability. Shapley Additive Explanations (SHAP) analysis under the feature-level fusion strategy indicated that appearance traits and chemical constituents made synergistic contributions to origin discrimination. The CatBoost model was further optimized using five multi-objective optimization algorithms, which improved the overall predictive performance for active component contents, with NSGA-II exhibiting a comparatively balanced overall advantage. The multimodal data fusion and multi-objective optimization framework developed in this study provides a new technical pathway for the origin traceability and quality evaluation of SM and offers a methodological reference for the intelligent analysis of multisource information in Chinese medicinal materials.