High accuracy prediction of Thai rice glycemic index using machine learning

This study investigated the effectiveness of machine learning (ML) models in estimating the glycemic index (GI) of Thai rice starches from their physicochemical characteristics. Three models, XGBoost, CatBoost and RandomForest, were employed on a dataset comprising various starch properties. All models yielded high R2 values exceeding 0.957, signifying their accuracy in GI prediction. Moreover, CatBoost exhibited superior performance with the highest test R2 value (0.977) and the lowest test root mean square error (RMSE) and mean absolute error (MAE) values (1.674 and 1.302, respectively). The CatBoost model was optimized to explore the impact of gelatinization parameters on GI prediction. Optimization revealed that maximizing the GI requires lower onset and peak temperatures, moderate conclusion temperature, and higher enthalpy, while minimizing the GI necessitates higher onset temperature and conclusion temperature, slightly higher peak temperature and slightly lower enthalpy. These findings provide valuable insights into tailoring starch properties to achieve desired GI levels in rice.