Prediction of Intracranial Aneurysm Risk using Machine Learning
Prediction of Intracranial Aneurysm Risk using Machine Learning
复制标题
DOI：
10.1038/s41598-020-63906-8
复制
发表时间：
2020-04-24
影响因子：
4.6
通讯作者：
Kim, Tackeun
中科院分区：
文献类型：
--
作者：
Heo, Jaehyuk;Park, Sang Jun;Kim, Tackeun
需要一种有效的方法来识别颅内动脉瘤(IA)的高危人群，以提供足够的放射学筛查指南并有效地分配医疗资源。我们利用国家索赔数据库和健康检查记录开发了一个用于诊断前IA预测的模型。来自韩国国家健康筛查计划的数据被用作几种机器学习算法的输入：Logistic回归(LR)、随机森林(RF)、可伸缩树增强系统(XGB)和深度神经网络(DNN)。使用不同于用于模型训练的测试数据，通过接收器操作特征曲线(AUROC)下的面积来评估算法性能。使用模型预测概率按风险升序对5个风险组进行分类。然后比较最低风险组和最高风险组之间的发病率比率。XGB模型预测IA风险最好(AUROC为0.765)，预测最低风险组的IA发生率(3.2)，而RF模型预测最高风险组的IA发生率(161.34)。在XGB、LR、DNN和RF模型中，最低风险组和最高风险组之间的发病率比分别为49.85、35.85、34.90和30.26。所开发的预测模型可为未来的IA筛查策略提供帮助。
An efficient method for identifying subjects at high risk of an intracranial aneurysm (IA) is warranted to provide adequate radiological screening guidelines and effectively allocate medical resources. We developed a model for pre-diagnosis IA prediction using a national claims database and health examination records. Data from the National Health Screening Program in Korea were utilized as input for several machine learning algorithms: logistic regression (LR), random forest (RF), scalable tree boosting system (XGB), and deep neural networks (DNN). Algorithm performance was evaluated through the area under the receiver operating characteristic curve (AUROC) using different test data from that employed for model training. Five risk groups were classified in ascending order of risk using model prediction probabilities. Incidence rate ratios between the lowest- and highest-risk groups were then compared. The XGB model produced the best IA risk prediction (AUROC of 0.765) and predicted the lowest IA incidence (3.20) in the lowest-risk group, whereas the RF model predicted the highest IA incidence (161.34) in the highest-risk group. The incidence rate ratios between the lowest- and highest-risk groups were 49.85, 35.85, 34.90, and 30.26 for the XGB, LR, DNN, and RF models, respectively. The developed prediction model can aid future IA screening strategies.