LSTM-NSGA-II framework for predicting and enhancing the quality of fermented kidney bean milk

Red kidney bean milk has high nutritional potential, but its application is limited by beany flavor and unstable fermentation performance. This study developed an integrated long short-term memory and non-dominated sorting genetic algorithm II (LSTM-NSGA-II) framework to model and optimize the fermentation of red kidney bean milk by Lactiplantibacillus plantarum PC4. Fermentation time, temperature, and inoculum level were optimized against five responses: DPPH, FRAP, total phenolic content (TPC), pH, and lactic acid bacteria (LAB) viability. The LSTM model showed strong predictive performance (test R2 = 0.9442-0.9958), and TOPSIS identified 6.46 h, 37.37 °C, and 1.99% inoculum as the best compromise condition. Experimental validation confirmed the reliability of the predicted optimum, with relative errors below 6%. Under optimized fermentation, physicochemical properties improved, and β-glucosidase activity increased markedly. Volatile profiling identified 96 compounds, with reduced off-flavor aldehydes, including hexanal and nonanal, and increased alcohols, acids, ketones, and esters associated with fruity, floral, roasted, and creamy notes. Fermentation also altered the sugar and organic acid profiles, reflected by L-lactic acid formation, decreased citric acid, slight sucrose reduction, and increased glucose. In vitro gastrointestinal tolerance results showed that L. plantarum PC4 retained measurable survival under acidic conditions. Sensory evaluation further indicated lower beany flavor intensity and higher fermented aroma and acidity in the fermented samples. These results demonstrate that the LSTM-NSGA-II framework is effective for precision fermentation design and for improving the functional and sensory quality of fermented kidney bean milk.

Text 125207.pdf - Published Version Restricted to Repository staff only Download (7MB) Official URL or Download Paper: https://linkinghub.elsevier.com/retrieve/pii/S2212...

Actions (login required)

View Item