CAJ | 학술논문

目的：比较BP神经网络模型和Cox比例风险模型在生存分析中的预测性能，进一步探讨BP神经网络模型在生存分析中的应用。方法：采用Monte Carlo 模拟数据集，如不同样本量、不同删失比例、不同协变量间关系及是否满足等比例风险假定的理论研究和胃癌根治术患者预后预测的实例分析，分别建立BP神经网络模型和Cox比例风险模型，最终使用一致性指数C对其预测性能进行比较。结果：当样本量为100、删失比例为60％、80％及样本量为300、删失比例为80％时，BP神经网络模型的预测性能高于Cox比例风险模型（P＜0．05）。协变量不满足等比例风险假定、协变量间存在三维交互作用和非线性关系时，BP神经网络模型的预测性能较Cox比例风险模型好（P＜0．05）。实例研究中发现，BP神经网络模型预测的一致性指数C（0．835）高于Cox比例风险模型（t配对＝4．311，P＜0．001）。结论：BP神经网络模型在生存分析的应用中对样本删失比例、是否满足PH假定、协变量间复杂交互作用和非线性关系具有非特异性，对资料限制较少，且预测一致性高，值得在生存分析中进一步推广应用。
목적：비교BP신경망락모형화Cox비례풍험모형재생존분석중적예측성능，진일보탐토BP신경망락모형재생존분석중적응용。방법：채용Monte Carlo 모의수거집，여불동양본량、불동산실비례、불동협변량간관계급시부만족등비례풍험가정적이론연구화위암근치술환자예후예측적실례분석，분별건립BP신경망락모형화Cox비례풍험모형，최종사용일치성지수C대기예측성능진행비교。결과：당양본량위100、산실비례위60％、80％급양본량위300、산실비례위80％시，BP신경망락모형적예측성능고우Cox비례풍험모형（P＜0．05）。협변량불만족등비례풍험가정、협변량간존재삼유교호작용화비선성관계시，BP신경망락모형적예측성능교Cox비례풍험모형호（P＜0．05）。실례연구중발현，BP신경망락모형예측적일치성지수C（0．835）고우Cox비례풍험모형（t배대＝4．311，P＜0．001）。결론：BP신경망락모형재생존분석적응용중대양본산실비례、시부만족PH가정、협변량간복잡교호작용화비선성관계구유비특이성，대자료한제교소，차예측일치성고，치득재생존분석중진일보추엄응용。
Aim:To compare their prediction performance of BP neural network model and Cox proportion hazard mod -el in survival analysis and to explore the superiority of BP neural network model in survival analysis .Methods: Monte Carlo was used to generate the data sets under the condition of different sample size , different degree of censoring , number of variable and interactions , non-linear effect , distinct distribution of covariate and proportional vs non-proportional hazard . Then BP neural network model and Cox model were built , and their prediction performance was compared using concord-ance index C.Results:In the research on simulation data sets , when the sample size of 100, proportion of censoring of 60%, 80%, and sample size of 300 , proportion of censoring of 80%, BP neural network model performed superior to Cox model(P<0.05).And when the covariates don′t meet PH assumption and had three-way interaction, non-linear effect, BP neural network performed superior to Cox model (P<0.05).In the real data, BP neural network model′s concordance index was 0.835, which performed superior to Cox model (tpaired =4.311,P<0.001).Conclusion:For the small sample size, high and the covariates don′t meet PH assumption and has three-way interaction, non-linear effect data sets , BP neu-ral network has better advantage than Cox model .It is worth to popularize further in survival analysis .