基于MRI的机器学习模型研究吸烟与大脑年龄的关系  

Study of the relationship between smoking and brain aging using machine learning model based on MRI

作　　者：高昕宇 张孟哲 韩少强 杨镇圭 汪卫建 许珂[1] 程敬亮[1] 张勇[1] Gao Xinyu;Zhang Mengzhe;Han Shaoqiang;Yang Zhengui;Wang Weijian;Xu Ke;Cheng Jingliang;Zhang Yong(Department of MR,First Affiliated Hospital of Zhengzhou University,Zhengzhou 450002,China)

机构地区：[1]郑州大学第一附属医院磁共振科,郑州450002

出　　处：《中华放射学杂志》2022年第12期1347-1351,共5页Chinese Journal of Radiology

基　　金：国家重点研发计划(SQ2018YFC130095);河南省科技攻关项目(172102310391);河南省重点研发与推广专项(科技攻关)支持项目(212102310712)。

摘　　要：目的通过基于MRI的机器学习模型来预测吸烟者及健康对照者大脑年龄,进一步探讨吸烟与大脑老化的关系。方法该研究为回顾性研究。数据集1为2014年8月至2017年10月郑州大学第一附属医院募集的男性吸烟者95名[年龄20~50(34±7)岁]和健康对照者49名[年龄20~50(33±7)岁]。数据集2为2010至2015年西南大学成人影像数据库的114名男性健康志愿者[年龄20~50(34±11)岁]。所有受检者均接受高分辨三维T1WI。基于数据集1和数据集2健康对照者的结构MR图像构建高斯过程回归(GPR)模型和支持向量机模型预测大脑年龄,并通过交叉验证法验证模型性能,计算预测大脑年龄与实际年龄间的平均绝对误差(MAE)、实际年龄和预测大脑年龄之间的相关性(r值),最终筛选出最佳模型。将最佳模型应用于吸烟者和健康对照者预测其大脑年龄。最后以年龄、教育年限及颅内总容积为协变量,通过一般线性模型比较吸烟者和健康对照者的大脑年龄差值(PAD)的差异。结果GPR模型预测大脑年龄(MAE=5.334,r=0.747)优于支持向量机模型(MAE=6.040,r=0.679)。GPR模型预测数据集1的吸烟者PAD值(2.19±6.64)高于数据集1的健康对照者(‒0.80±8.94),差异有统计学意义(F=8.52,P=0.004)。结论基于MRI的GPR模型预测吸烟者及健康对照者大脑年龄性能较好,吸烟者PAD值增加,进一步表明吸烟会加速大脑老化。Objective To explore the value of machine learning models based on MRI predict the brain age of smokers and healthy controls,and further to explore the relationship between smoking and brain aging.Methods This was a retrospective study.Dataset 1 consisted of 95 male smokers[20-50(34±7)years old]and 49 healthy controls[20-50(33±7)years old]recruited from August 2014 to October 2017 in First Affiliated Hospital of Zhengzhou University.Dataset 2 contained 114 healthy male volunteers[20-50(34±11)years old]from the Southwestern University Adult Imaging Database from 2010 to 2015.All subjects underwent high-resolution 3D T1WI scan.Gaussian process regression(GPR)model and support vector machine model were constructed to predict brain age based on structural MR images of healthy controls in dataset 1 and dataset 2.After the performance of the model was verified by the cross-validation method,the mean absolute error(MAE)between the predicted brain age and the actual age and the correlation(r-value)between the actual age and the predicted brain age were calculated,and the best model was finally selected.The best models were applied to smokers and healthy controls to predict brain age.Finally,a general linear model was used to compare the differences in brain-predicted age difference(PAD)between smokers and healthy controls with age,taking years of education and total intracranial volume as covariates.Result The performance of GPR model(MAE=5.334,r=0.747)in predicting brain age was better than support vector machine model(MAE=6.040,r=0.679).The GPR model predicted that PAD value of smokers in dataset 1(2.19±6.64)was higher than that of healthy controls in dataset 1(-0.80±8.94),and the difference was statistically significant(F=8.52,P=0.004).Conclusion GPR model based MRI has better performance in predicting brain age in smokers and healthy controls,and smokers show increased PAD values,further indicating that smoking accelerates brain aging.

关 键 词：吸烟 磁共振成像 大脑年龄 机器学习 

分 类 号：R445.2[医药卫生—影像医学与核医学]