AI-Driven Nutritional Profiling of Goji Berry (Lycium barbarum L.) Through Multi-Source Data Integration and Explainable Machine Learning

Published goji berry composition data differ in analytical basis and sampling design. We assembled the Goji Berry Multi-Source Nutritional Database (GB-MND; 10,568 records) and reanalyzed two published datasets separately: 78 dry-weight samples with eight variables and 186 powder samples with 44 metabolites. PCA and clustering described composition patterns; four machine learning models classified province labels within each dataset using repeated nested cross-validation. The highest macro-F1 values were 0.732 and 0.982, respectively. In the first dataset, region-wise evaluation reduced macro-F1 to 0.226–0.287; the second dataset had one site per province, preventing a separate assessment of transfer to new sites of represented classes. Out-of-fold interpretation and sensitivity analyses retained four and 11 model-associated components. Exploratory comparisons with two independent studies gave PC1 loading cosines of 0.996 and 0.910, with wide bootstrap intervals of 0.109–0.995 and 0.039–0.957. Polysaccharide median concentrations were close to the development dataset in both comparisons, while several carotenoid concentrations differed substantially. The reanalysis identifies recurring compositional associations and shows how their interpretation changes with sampling and analytical context. GB-MND provides source-linked data and reproducible analyses for developing harmonized goji composition panels.