可解释的生存分析机器学习：AutoScore-Survival方法

版权申诉

111 浏览量更新于2024-07-06 收藏 662KB PDF 举报

"AutoScore-Survival 是一个用于开发可解释的机器学习时间到事件分数的框架，专门处理右删失生存数据。该框架由Feng Xie、Yilin Ning、Han Yuan等人提出，结合了Duke-NUS Medical School、Duke University等多个机构的研究成果。AutoScore-Survival的目标是提供一种透明且实用的方法，用于预测和理解与生存时间相关的复杂模式，尤其在医疗保健研究中评估事件发生的时间结果。正文: AutoScore-Survival是一个创新的统计工具，它利用机器学习方法来构建对生存分析具有高解释性的评分系统。在医学研究中，尤其是在临床试验和流行病学研究中，时间到事件分析（如生存分析）是评估患者预后或疾病进展的关键。然而，传统的生存模型如Cox比例风险模型往往难以解释复杂的协变量关系，而机器学习方法虽然可以捕捉这些复杂性，但通常缺乏可解释性。该框架的独特之处在于它能够处理右删失数据，这是生存数据分析中常见的问题。右删失意味着对于某些个体，我们只知道他们没有在研究期间发生事件，但不知道确切的发生时间。AutoScore-Survival通过集成事件时间分数和机器学习技术，解决了这一挑战，提供了预测个体生存时间的分数。在论文中，作者们详细介绍了AutoScore-Survival的构建过程，包括特征选择、模型训练和验证。他们可能使用了各种算法，如随机森林、梯度提升机等，这些算法能够在大量变量中找到最相关的特征，并建立预测模型。此外，为了确保模型的解释性，AutoScore-Survival还可能引入了特征重要性评估，使得每个特征对最终分数的贡献清晰可见。 AutoScore-Survival的应用场景广泛，包括但不限于心血管疾病、癌症研究、慢性疾病管理等领域。通过创建直观的评分系统，医生和研究人员可以更准确地预测患者的生存概率，制定个性化的治疗策略，从而改善患者的健康结局。此外，由于AutoScore-Survival强调可解释性，它可以帮助医学界更好地理解和接受基于机器学习的预测模型。这对于推动临床决策支持系统的实施至关重要，因为医生需要理解模型背后的逻辑才能信任并应用这些工具。总结来说，AutoScore-Survival是机器学习与生存分析领域的交汇点，旨在提供一种既强大又可解释的工具，用于处理医疗研究中的时间到事件预测问题。通过结合统计学和计算科学，这个框架为医学研究和临床实践带来了新的可能性，促进了对生存数据更深入的理解和利用。"

Model selection under the intermediate performance evaluation at Module 4

The validation set is used for the intermediate performance evaluation. We use a survival

parsimony plot to visualize the change in model performance with an increasing number of

variables, which helps us select a model that balances prediction accuracy and parsimony. For

time-to-event outcomes, the time-dependent area under the curve or AUC(t)

is applied to

measure model performance, which is an extension of the commonly used area under the

curve (AUC) for measuring predictive accuracy of a score when studying binary outcomes.

We chose the AUC(t) defined by cumulative sensitivity and dynamic specificity (C/D) as

recommended by a comprehensive review

, as this definition has more clinical relevance and

commonly been used by clinical application

. This AUC(t) is introduced as a function of

time, estimated through the Kaplan-Meier estimator of survival function

, to characterize

how well the score can distinguish between subjects who had an event <= t and remained

event-free at time t. To obtain a single overall performance metric in the parsimony plot, we

derived the integrated AUC (iAUC), a weighted average of AUC(t)

over a follow-up period

(i.e., from Day 1 to Day 90), summarizing the overall discrimination ability of the time-to-

event score (see eTextbox 2 in Appendix for detail).

Final predictive performance evaluation at Module 6

We evaluate the final time-to-event score in the test set using multiple performance metrics.

In addition to iAUC and AUC(t), we used the Harrell's concordance index (C-index)

36, 37

which is the proportion of concordant pairs (i.e., when the observation with a longer survival

time has a larger time-to-event score) in all pairs formed in the test set. Thus, the C-index is

able to summarize risk, event occurrence, and survival time in a single number to distinguish

between well-behaved scores and quasi-random ones

Algorithm Implementation and empirical validations

We implemented the AutoScore-Survival framework as an R package

. Given a new dataset

with time-to-event outcomes and baseline covariates, the AutoScore-Survival package

provides a pipeline of functions to split data and implement the six modules to generate the

final scores that require minimal manipulation from users.

We demonstrated our AutoScore-Survival algorithm using the same dataset as our previous

paper

, including 44,918 ICU admissions from Beth Israel Deaconess Medical Center

(BIDMC)

with 24 available variables of demographic information, vital signs, and lab tests

at baseline (t=0). The survival status and the date of death were additionally obtained from

the database to derive the 90-day survival as the primary outcome. The baseline

characteristics of the dataset were described through univariable and multivariable Cox

regression. The Kaplan-Meier survival curves were generated for different risk groups

stratified by the scores and compared through the log-rank test. Furthermore, we compare the

predicted and actual survival probabilities at different time points. To evaluate the

performance of AutoScore-Survival, we compared it with several standard time-to-event

prediction models. We considered the Cox model with (i) all variables and that with variables

selected using (ii) stepwise and (iii) LASSO

approaches. The stepwise variable selection

used AIC and considered both directions in each step. The regularization rate of LASSO was

optimized through 10-fold cross-validation). We also built an RSF using all variables, with

剩余21页未读，继续阅读

易小侠

粉丝: 6605
资源: 9万+

可解释的生存分析机器学习：AutoScore-Survival方法

AutoScore不平衡：一种可解释的机器学习工具，用于开发具有罕见事件数据的临床评分_AutoScore-Imbalance

AutoScore.rar

matlab倒频谱代码-SWD_AUTOSCORE:SWD_AUTOSCORE

matlab倒频谱代码-SWD-Automatic-Identification:用于自动检测syngap1大鼠SWD的Matlab脚本

数学建模学习资料 神经网络算法 参考资料-Matlab 共26页.pptx

happybirthday2 升级版生日祝福密码0000(7).zip

ssm框架Java项目源码-基于web技术的税务门户网站的实现+vue毕设-大作业.zip

php7.4.33镜像7z压缩包

ssm框架Java项目源码-基于java的珠宝购物网站系统的建设+jsp毕设-大作业.zip

基于java的高校大学生党建系统设计与实现.docx

最新资源

数学建模学习资料神经网络算法参考资料-Matlab 共26页.pptx