Multi-source fusion of NIR and UV spectroscopy with refractometry for pooled prediction of total flavonoids, total polyphenols and solids content in herbal water extracts.
Abstract
Rapid quantification of multiple quality indicators in compositionally distinct herbal formulas calls for predictive approaches that can accommodate pooled samples within a shared modeling framework. To address this limitation, the present study integrates multi-source information from near-infrared (NIR) spectroscopy, ultraviolet (UV) spectroscopy, and refractive index (RI) measurements to develop one pooled model for each of three indicators: total flavonoids, total polyphenols and total solids content in aqueous extracts of six TCM formulas. Six TCM formulas with markedly distinct herbal compositions were selected, and 216 samples were prepared through a gradient extraction protocol. Under a unified SGS + FD pretreatment, three fusion architectures, five feature selection algorithms, and 16 machine learning models were systematically compared. Low-level fusion (LowFusion) performed best for flavonoids and polyphenols before feature selection, whereas mid-level fusion performed slightly better for solids content; the optimal feature selection algorithm was analyte-dependent, and the selected wavelength regions were consistent with plausible electronic-transition and vibrational assignments. Validation using 18 independent samples from new batches of the same six formulas showed strong predictive performance for flavonoids and solids content. The polyphenol model displayed systematic underestimation in the high-concentration range. These results support pooled multi-formula modeling and independent-batch prediction within the investigated formula set, providing a basis for rapid multi-indicator quality assessment without developing a separate model for each formula.