目的 旨在将传统藏医药性理论与现代机器学习方法相结合，构建基于营养素特征的藏医饮食甘味预测模型，为藏医药性理论的现代化解读提供科学依据。方法 系统梳理藏医药经典文献与现代食物营养数据，构建包含786种食物的藏医饮食数据库。通过递归特征消除和五折交叉验证（recursive feature elimination and five-fold cross-validation，FRECV），筛选关键营养特征变量，并基于7种机器学习算法构建甘味预测模型。采用CatBoost最佳模型联合SHAP解释框架分析特征重要性与模型预测机制，验证模型的解释性与个体化预测能力。结果 从29种营养素中最终筛选出了8个核心特征，其中矿物质元素（铁、钙、钠、磷、钾）占比62.5%，表明矿物质在甘味食物分类中具有重要作用。CatBoost模型在多项性能指标上表现最优，测试集曲线下面积（area under curve，AUC）值达0.78，准确率0.83，召回率1.00。SHAP分析显示钙和铁为影响甘味分类的最关键因素，并揭示矿物质特征间存在协同与拮抗效应。决策路径分析表明，模型具备良好的特征分层识别与个体化预测能力。结论 首次通过机器学习方法系统解析藏医甘味食物的营养特征，验证矿物质元素在“土水生甘”理论中的关键作用，为藏医饮食“味性化味”理论提供现代科学依据。所构建的甘味预测模型具备良好的泛化性能和可解释性，为高原地区膳食优化、藏医药食同源功能食品设计提供理论基础和技术支撑。未来研究将结合多模态数据和临床试验，进一步提升模型的实用性与精准干预能力。

Objective This study aims to combine traditional Tibetan medicine property theory with modern machine learning methods to construct a Tibetan dietary sweet taste prediction model based on nutritional characteristics, providing scientific evidence for the scientific modernization of Tibetan medicine property theory. Methods A Tibetan dietary database containing 786 foods was constructed by systematically reviewing classic Tibetan medicine literature and modern food nutrition data. Key nutritional feature variables were screened through recursive feature elimination and five-fold cross-validation (RFECV), and sweet taste prediction models were constructed based on seven machine learning algorithms. The feature importance and model prediction mechanism were analyzed using the CatBoost optimal model combined with the SHAP interpretation framework to verify the interpretability and individualized prediction capability of the model. Results Eight core features were ultimately selected from 29 nutrients, with mineral elements (iron, calcium, sodium, phosphorus, potassium) accounting for 62.5%, indicating that minerals play an important role in the classification of sweet-tasting foods. The CatBoost model performed best on multiple performance metrics, with an area under curve (AUC) value of 0.78, accuracy of 0.83, and recall of 1.00 on the test set. SHAP analysis showed that calcium and iron were the most critical factors affecting sweet taste classification, and revealed synergistic and antagonistic effects between mineral features. Decision path analysis indicated that the model has good feature hierarchical recognition and individualized prediction capabilities. Conclusion This study systematically analyzed the nutritional characteristics of Tibetan sweet-tasting foods through machine learning methods for the first time, verifying the key role of mineral elements in the “earth and water produce sweetness” theory, providing modern scientific evidence for the Tibetan dietary “flavor-nature transforms flavor” theory. The constructed sweet taste prediction model has good generalization performance and interpretability, providing a theoretical foundation and technical support for dietary optimization in plateau regions and the design of functional foods with the same source of medicine and food in Tibetan medicine. Future research will combine multimodal data and clinical experiments to further enhance the practicality and precision intervention capabilities of the model.

TP18;R285.1

国家社科冷门绝学研究专项学术团队项目（23VJXT026）；国家自然科学基金青年项目（62206146）；青海省科技厅青年项目（2023-ZJ-991Q）