基于SHAP的三阴性乳腺癌可解释预测模型的建立  

The establishment of the interpretable predictive model for triple negative breast cancer based on SHAP

作　　者：刘孟昕 葛敏[1] 王世威[1] 陆欢 Liu Mengxin

机构地区：[1]浙江中医药大学附属第一医院,310006

出　　处：《浙江临床医学》2024年第4期487-489,共3页Zhejiang Clinical Medical Journal

基　　金：浙江省公益技术应用研究资助项目(LGF21H180003);浙江省中医药科学研究基金资助项目(2022ZB132)。

摘　　要：目的为三阴性乳腺癌患者构建一种能够同时获得良好效果的、可解释的预测模型。方法回顾性分析136例乳腺癌患者的临床特征和多序列多参数核磁共振成像,其中三阴性乳腺癌23例,非三阴性乳腺癌113例。通过勾画提取影像组学特征进行筛选并构建模型,最后结合放射组学特征和独立的临床图像特征,构建机器学习框架。此外,还采用为实现个性化临床决策支持提供个性化评估和解释的SHAP模型可解释器。结果经过影像组学特征筛选,11个特征参与计算影像组学评分,其在训练集与测试集的AUC为0.898、0.803。将其与临床模型结合,使预测精度进一步提高。结论多模式可解释预测模型可能会帮助临床医师更准确、更迅速识别三阴性乳腺癌风险,及时、准确为患者治疗。Objective To construct an interpretable prediction model for triple negative breast cancer patients,which can simultaneously achieve good prediction and interpretation capabilities.Methods The clinical features and multi sequence and multi parameter MRI images of 136 patients with breast cancer were retrospectively analyzed,including 23 cases of triple negative breast cancer and 113 cases of non triple negative breast cancer.After screening and constructing the model by sketching and extracting the radiomic features,the machine learning framework was constructed by combining the radiomic features and independent clinical image features.In addition,the SHAP(Sharpley Additive exPlanning)model interpreter was used to provide personalized evaluation and interpretation to achieve personalized clinical decision support.Results After screening the omics features,11 radiomic features were involved in the calculation of the radiomic score,and their AUC in the training set and the test set were 0.898 and 0.803.The prediction accuracy was further improved by combining with the clinical model.Conclusion The multimodal interpretable prediction model may help clinicians identify triple negative breast cancer risk patients more accurately and quickly,and provide timely and accurate treatment for patients.

关 键 词：三阴性乳腺癌 磁共振 影像组学 SHAP算法 

分 类 号：R737.9[医药卫生—肿瘤]