从SHAP到概率——可解释性机器学习在糖尿病视网膜病变靶向脂质组学研究中的应用

From SHAP to Probability——An Interpretable Machine Learning Framework for Targeted Lipidomics Study on Diabetic Retinopathy

目的基于可解释性机器学习算法构建糖尿病视网膜病病变(diabetic retinopathy,DR)的早期识别模型,并探讨SHAP(SHapley Additive exPlanations)在脂质组学数据中的应用。方法基于本项目组的DR靶向脂质组学数据,通过可解释性机器学习的方法进行特征筛选;在建立糖尿病视网膜病变的早期识别模型后,通过全局、特征和个体三个层面对模型进行解释,并将SHAP值转换成概率以增强可解释的能力。结果本研究筛选出了5种内源性脂质代谢物,构建了一个性能较为优秀的糖尿病视网膜病变的早期识别模型,并成功使用SHAP及概率解锁了模型。结论脂质代谢物质可以应用于糖尿病视网膜病变的早期识别;SHAP在进行黑盒模型的解锁时表现出色,且有较高的实践应用价值。

Objective To Construct an early identification model for diabetic retinopathy(DR)based on machine learning algorithms and explore the application of SHAP to lipidomics data.Methods Based on the DR quantitative lipidomics data from our project group,feature screening was performed by an interpretable machine learning approach;after building a model for early identification of diabetic retinopathy,the model was interpreted at three levels:global,feature and individual,and SHAP values were converted into probabilities to enhance interpretability.Results Five endogenous lipid metabolites were screened,and a model with better performance for early identification of diabetic retinopathy was constructed and successfully unlocked using SHAP and probability.Conclusion Lipid metabolites can be used for early identification of diabetic retinopathy;SHAP performs well in performing unlocking of black box models and has high practical application.