机器学习预测渤海中麻痹性贝类毒素及其膳食风险评估

Prediction of paralytic shellfish poisoning in the Bohai Sea by machine learning and its dietary risk assessment

中文关键词:  麻痹性贝类毒素  重金属  预测  特征选择  机器学习

英文关键词:paralytic shellfish poisoning  multi metallic elements  prediction  feature selection  machine learning

中文摘要:

      目的 实现渤海中麻痹性贝类毒素(paralytic shellfish poisoning, PSP)预测, 分析与PSP产生有关的金属污染物元素, 并完成PSP的膳食风险评估。方法 本研究对2018—2024年渤海领域石房蛤毒素(saxitoxin, STX)、膝沟藻毒素(gonyautoxins2/3/4, GTX2、GTX3、GTX4)、新石房蛤毒素(neosaxitoxin, neoSTX)共5种PSP检出率相对较高的样本进行检测分析, 使用最大信息系数(maximal information coefficient, MIC)计算与贝体内金属多元素含量的相关性、冗余性和互补性, 并基于此进行金属元素特征选择, 对筛选后的特征子集使用5种经典机器学习模型实现PSP含量预测。实验采用十折交叉验证结果, 用配对t检验分析, 用95%置信区间(confidence interval, CI)比较数据集的预测准确率。最后采用点评估法针对5个年龄段的人群的环渤海领域毛蚶和海虹两种贝类膳食风险进行的评估。结果 使用特征选择方法可有效筛选与PSP含量有关的金属多元素危害因子, 随机森林(random forest, RF)模型使用特征子集预测膝沟藻毒素中GTX3的平均绝对误差(mean absolute error, MAE)、均方根误差(root mean square error, RMSE)下降最大, 分别降低了8.0% (95%CI: –0.0821~–0.0779, P<0.05)和9.1% (95%CI: –0.0997~–0.0823, P<0.05), RF模型在预测中具有更低的预测误差, MAE、RMSE分别为0.147和0.183。5个年龄段的人群食用环渤海领域毛蚶和海虹贝类膳食风险均未超过限值。结论 本研究提出的融合MIC与冗余-互补性评估的动态特征选择方法, 能有效挖掘金属多元素间的协同与拮抗作用, 显著提升PSP预测精度, 同时5个年龄段人群相关贝类的膳食风险均处于安全水平, 为渤海贝类食品安全管控提供了科学支撑与技术参考。

英文摘要:

      Objective To predict the paralytic shellfish poisoning (PSP) in the Bohai Sea, explore the correlation between foodborne multi metallic elements contaminants and the levels of PSP, and complete the dietary risk assessment of PSP. Methods This study analyzed samples with relatively high detection rates of 5 kinds of PSP, including saxitoxin (STX), gonyautoxins2/3/4 (GTX2, GTX3, GTX4) and neosaxitoxin (neoSTX), collected in the Bohai Sea from 2018 to 2024. The maximal information coefficient (MIC) was used to analyze the correlation, redundancy and complementarity with the multi metallic element content in bivalves. Based on this, feature selection for multi metallic elements was conducted, and 5 kinds of classical machine learning models to predict PSP levels was utilized. The experiments employed 10-fold cross-validation, paired t-tests and compared the prediction accuracy using a 95% confidence interval (CI). Finally, the point assessment method was used to evaluate the dietary risks of 2 kinds of shellfish, Manila clams and mussels, in the Bohai Sea region for 5 different age groups. Results The feature selection method effectively identified multi metallic element hazard factors related to PSP content. The random forest (RF) model showed the largest reductions in mean absolute error (MAE) and root mean square error (RMSE) for predicting GTX3 of gonyautoxin, with reductions of 8.0% (95%CI: –0.0821 to –0.0779, P<0.05) and 9.1% (95%CI: –0.0997 to –0.0823, P<0.05) respectively. The RF model exhibited lower predictive errors, with MAE and RMSE values of 0.147 and 0.183 respectively. The dietary risks of consuming Manila clams and mussels from the Bohai Sea area were acceptable for people in all 5 age groups. Conclusion The dynamic feature selection method proposed in this study, which integrates MIC and redundancy-complementarity evaluation, can effectively uncover synergistic and antagonistic effects among multiple metal elements, significantly improving PSP prediction accuracy. Meanwhile, the dietary risks of relevant bivalves for all 5 age groups remain within safe levels, providing scientific support and technical reference for the safety management of bivalves in the Bohai Sea.