Combining H2O auto-ML and traditional machine learning for the resource evaluation on non-medicinal parts of plants: taking Coptis deltoidea as an example.

The resource evaluation and development of non-medicinal parts of traditional medicinal herbs have significant implications. However, the traditional detection methods and analytical techniques are hindered by their complex operations and time-consuming characteristics, which limits their ability for rapid evaluation. Hence, a H2O autoML approach was employed. This study analyzed a Fourier-transform near-infrared spectroscopy (FT-NIR) dataset of 580 samples of Coptis deltoidea C.Y. Cheng et Hsiao (including root, stem, and leaf) from different origins and harvest years. Moreover, multiplicative scatter correction (MSC) combined with second derivative (SD) smoothing, and spectral standard deviation were applied. Alkaloid content in the non-pharmaceutical parts was measured by Ultra high-performance liquid chromatography (UHPLC). Results showed that H2O AutoML and K-nearest neighbor (KNN) achieved high accuracy among the different classification models. Moreover, in regression analysis, H2O AutoML outperformed conventional models with Residual prediction deviation (RPD) values in 1.514-2.309. Shapley additive explanations (SHAP) analysis indicated that absorptions at 7200-7000 cm-1, likely associated with hydroxyl compounds and free water, contributed negatively to alkaloids prediction. Finally, correlation analysis revealed that potassium (K) and phosphorus (P) were key elements influencing berberine accumulation. The findings provide a rapid and reliable analytical method for the resource evaluation and development of abundant non-medicinal plants.