循环炎症相关因子神经网络模型预测脑卒中后抑郁发生风险的效能分析
A neural network model based on circulating inflammation-related factors for risk of PSD:construction and prediction efficiency analysis
李凤玲 1杨学 1陈海燕1
作者信息
- 1. 430070 武汉科技大学附属老年病医院脑科中心
- 折叠
摘要
目的 探讨基于神经网络算法构建循环炎症相关因子对脑卒中后抑郁(post-stroke depression,PSD)发生风险预测模型.方法 前瞻性选取2021年3月至2024年3月武汉科技大学附属老年病医院脑科中心脑卒中就诊的患者260例,其中训练集208例(80％)和验证集52例(20％),根据脑卒中后1个月内PSD发生情况将训练集脑卒中患者分为PSD组(62例)和非PSD组(146例).通过训练集筛选影响PSD发生风险的预测因素,在训练集中基于多因素logistic和神经网络算法分别构建PSD发生风险预测模型,比较2个预测模型的预测效能,同时在验证集进行验证.结果 本研究脑卒中后1个月内发生PSD 76例(29.23％),其中训练集62例,验证集14例.PSD组C反应蛋白(C-reactive protein,CRP)、纤维蛋白原(fibrinogen,FIB)、白细胞介素(interleukins,IL)-6、IL-1β、肿瘤坏死因子 α(tumor necrosis factor-α,TNF-α)、IL-18、中性粒细胞与淋巴细胞比值(neutrophil and lymphocyte ratio,NLR)明显高于非PSD组,差异有统计学意义(P＜0.01).多因素logistic回归分析显示,CRP(OR=1.494,95％CI:1.239～1.802)、FIB(OR=1.924,95％CI:1.191～3.109)、IL-6(OR=1.128,95％CI:1.001～1.272)、TNF-α(OR=1.051,95％CI:1.010～1.093)、IL-1β(OR=1.096,95％CI:1.006～1.194)、IL-18(OR=1.019,95％CI:1.002～1.036)、NLR(OR=1.873,95％CI:1.027～3.418)为 PSD 发生风险的危险因素(P＜0.05,P＜0.01).ROC曲线结果显示,神经网络算法的预测模型的曲线下面积明显高于多因素logistic回归分析模型(0.931 vs 0.855,Z=3.448,P＜0.05),且基于验证集评估,神经网络模型的准确性明显高于多因素logistic模型(92.31％vs 75.00％,P＜0.05).结论 循环炎症相关因子CRP、FIB、IL-6、IL-1β、TNF-α、IL-18、NLR与PSD发生风险有关,基于神经网络算法构建的循环炎症相关因子预测模型能更有效预测PSD发生风险.
Abstract
Objective To construct a risk prediction model for post-stroke depression based on the neural network algorithm.Methods A prospective study was conducted on 260 stroke patients admitted in our center from March 2021 to March 2024.They were randomly divided into a train-ing set(80％,208 cases)and a verification set(20％,52 cases).According to the occurrence of post-stroke depression within 1 month after stroke,the training set was assigned into post-stroke depression group(62 cases)and non-post-stroke depression group(146 cases).The predictive fac-tors for post-stroke depression occurrence were screened through the training set,and the risk prediction models for post-stroke depression occurrence were constructed with multivariate logis-tic and neural network algorithms in the training set.The prediction efficiency of the two predic-tion models was compared and verified in the verification set.Results Within 1 month after stroke,76 cases(29.23％)experienced post-stroke depression(62 cases in training set and 14 in the validation set).Based on the data in the training set,the levels of CRP,FIB,IL-6,IL-lβ,TNF-αand IL-18,and neutrophil and lymphocyte ratio(NLR)were significant higher in the post-stroke depression group than the non-post-stroke depression group(P＜0.01).Multivariate logistic re-gression analysis showed that CRP(OR=1.494,95％CI:1.239-1.802),FIB(OR=1.924,95％CI:1.191-3.109),IL-6(OR=1.128,95％CI:1.001-1.272),TNF-α(OR=1.051,95％CI:1.010-1.093),IL-1β(OR=1.096,95％CI:1.006-1.194),IL-18(OR=1.019,95％CI:1.002-1.036),and NLR(OR=1.873,95％CI:1.027-3.418)were risk factors for post-stroke depression(P＜0.05,P＜0.01).ROC curve analysis indicated that the AUC value of the predictive model of the neural network algorithm was higher than that of the model of multivariate logistic regression(0.931 vs 0.855,Z=3.448,P＜0.05).Based on the validation set,the former model also had bet-ter accuracy than the latter one(92.31％vs 75.00％,P＜0.05).Conclusion Circulating inflam-matory factors CRP,FIB,IL-6,IL-1β,TNF-α and IL-18,and NLR are related to the risk of post-stroke depression.The prediction model based on above factors combined with neural network al-gorithm can more effectively predict the risk of post-stroke depression.
关键词
卒中/抑郁/比例危险度模型/神经网络模型Key words
stroke/depression/proportional hazards models/neural network model引用本文复制引用
出版年
2025
NETL
NSTL
万方数据