Rapid adulteration detection of herbal powder using AI-driven multiple digital fingerprints: A case of Bajiaohuixiang (Illicium verum Hook. f.)

Herbal powder, a common dosage form in traditional Chinese medicine, requires rigorous authenticity assessment to ensure medicinal quality and therapeutic efficacy. In this study, we proposed an artificial intelligence (AI) assisted digital fingerprint strategy for rapid detection of herbal powder adulteration, using Bajiaohuixiang (BJHX) as a case. The intelligent algorithms combined with digital image recognition (DIR) and near-infrared spectroscopy (NIRS) techniques achieved effective discrimination in both binary and seven-class classification of adulteration. The optimized bidirectional long short-term memory network delivered good performance, which achieved accuracies of 96.11% and 80.00% using the DIR dataset, and 94.44% and 88.10% using the NIRS dataset, for binary and seven-class classification, respectively. Multimodal data fusion significantly enhanced performance, yielding classification accuracies of 97.22% and 91.90% for the two tasks, respectively, and a quantitative model with coefficient of determination of 0.8807, root mean square error of 0.0560, and mean absolute error of 0.0443. The strategy's robustness was confirmed through external validation, where only one out of multiple commercial and laboratory-prepared samples was flagged as potentially adulterated. Furthermore, greenness assessment using Analytical Greenness Calculator (AGREE, 0.92), Analytical Greenness Metric of Sample Preparation (AGREEprep, 0.90), and Blue Applicability Grade Index (BAGI, 80) frameworks confirmed its environmental sustainability and practical applicability. Overall, this work highlights the significant potential of integrating AI with DIR and NIRS as a rapid, accurate, and green tool for herbal powder authentication. • AI-assisted DIR and NIRS enabled rapid detection of BJHX powder adulteration. • Adapted Bi-LSTM achieved superior performance in classification and quantification. • Feature-level data fusion significantly enhanced the model ability. • External validation realized accurate prediction of prepared and market samples. • The established method adhered to green analytical chemistry principles.