点击率预测问题建模_使用预测建模技术预测住院率

点击率预测问题建模

机器学习(Machine Learning)

Hospital readmissions, particularly unplanned hospital readmissions are costly to the health care industry. Centers for Medicare and Medicaid Services (CMS) reporting an annual $17 billion in health care spending as a result of hospital readmissions. CMS outlined chronic conditions with a high risk of frequent hospitalizations/readmissions as part of the 2010 Hospital Readmission Reduction Program.

^ h ospital再住院，特别是计划外再入院是昂贵的医疗保健行业。 医疗保险和医疗补助服务中心(CMS)报告，由于医院再次入院，每年在医疗保健方面的支出为170亿美元。 CMS概述了慢性病，这是经常住院/再入院的高风险，是2010年医院再入院减少计划的一部分。

The rise of EHR technology aided in tracking and capturing of patient data, particularly patients that fell into the category of “high risk” for readmission. Considering the CMS’s outline of chronic conditions and the use of EHR technology, the stage was set for the question of the possibility to predict readmissions.

EHR技术的兴起有助于跟踪和捕获患者数据，特别是那些属于再次入院“高风险”类别的患者。 考虑到CMS的慢性病概述和EHR技术的使用，为预测再入院的可能性问题设定了阶段。

The accuracy of predicting readmissions varies across predictive modeling techniques. Predicting readmissions has since evolved from the standard hospital administrative risk assessments such as LACE to machine learning techniques. Machine learning techniques varying from the baseline model of Logistic Regression to more advanced and complicated models such as Deep Neural Networks.

预测再入院的准确性因预测建模技术而异。 此后，重新入学的预测已从标准的医院行政风险评估(例如LACE)演变为机器学习技术。 机器学习技术从Logistic回归的基线模型到更高级和复杂的模型(如深度神经网络)不等。

The objective of this literature review is to discuss the most commonly used predictive modeling strategies and their accuracy in predicting hospital readmissions.

这篇文献综述的目的是讨论最常用的预测建模策略及其在预测住院率方面的准确性。

The databases used in the search for relevant articles were: Adelphi University Library Online Database (OneSearch), PubMed, Google Scholar, ScienceDirect, Wiley Online Library as well as British Medical Journal, New England Medical Journal, JAMA Network, Springer Link, and PLOS One. 35 articles were discovered matching the keywords, however after reviewing the type of article (ie, relevance, year article was written- articles written > 20 years were excluded, articles that were not peer-reviewed were also excluded) the articles were then narrowed down to 27.

搜索相关文章时使用的数据库为：阿德菲大学图书馆在线数据库(OneSearch)，PubMed，Google Scholar，ScienceDirect，Wiley在线图书馆以及《英国医学杂志》，《新英格兰医学杂志》，JAMA Network，Springer Link和PLOS一。 发现35个与关键字匹配的文章，但是在审查了文章类型之后(即，相关性，撰写年份文章-排除了写作> 20年的文章，也排除了未经同行评审的文章)，然后缩小了文章的范围至27。

The overall consensus was deep neural networks outperformed traditional predictive modeling techniques such as LACE and machine learning (Logistic Regression). There is still much to be learned to improve predictive modeling. Health care is complex and the system is not one size fits all, there is a need to take into consideration patient health conditions and other factors that influence health care outcomes.

总体共识是，深度神经网络优于传统的预测建模技术，例如LACE和机器学习(逻辑回归)。 改善预测模型仍有很多知识要学习。 卫生保健很复杂，而且该系统并不是一个适合所有人的规模，因此需要考虑患者的健康状况和其他影响卫生保健结果的因素。

Factors such as social and economics play a role in health care outcomes. These factors have been known to typically be excluded in predictive modeling techniques/strategies, however, research suggests that such factors have an impact on health care outcomes and predictive modeling techniques.

社会和经济因素在医疗保健结果中发挥作用。 众所周知，这些因素通常在预测建模技术/策略中不包括在内，但是研究表明，这些因素会对医疗保健结果和预测建模技术产生影响。

The future of predictive modeling in health care is a system that is able to take into consideration the patient as a whole, while factoring patient social and economic barriers/status in order to improve patient predictions.

卫生保健中的预测建模的未来是一个能够考虑到患者整体的系统，同时考虑到患者的社会和经济障碍/状况以改善患者预测。

介绍 (Introduction)

Hospital readmissions continue to greatly account for health care spending in the United States. Many of the readmissions, according to Centers for Medicare and Medicaid Services (CMS), are considered avoidable and preventable. Frequent hospital readmissions indicate a lack of quality care, poor discharge planning, as well as a wealth of other social and economic factors which may have an effect on readmissions rates.

住院再入院继续在美国占医疗保健支出的很大比例。 根据医疗保险和医疗补助服务中心(CMS)的说法，许多重新入院被认为是可以避免和预防的。 频繁的医院再入院表明缺乏优质的医疗服务，出院计划不佳以及其他许多可能影响再入院率的社会和经济因素。

CMS created the Hospital Readmission Reduction Program (HRRP) in 2010 in an effort to reduce frequent hospital readmissions. The program focuses on reducing Medicare payments if a patient is readmitted to a hospital within 30 days from discharge date and the readmission is considered preventable or avoidable. Medicare payments are to be reduced by up to 3%.

CMS于2010年制定了减少医院再入院计划(HRRP)，以减少频繁的医院再入院。 该计划的重点是，如果患者在出院之日起30天内再次入院并且被认为是可以预防或避免的，则应减少医疗保险费用。 医疗保险付款最多可减少3％。

CMS outlined health conditions considered high risk for readmission, calling for improved coordination of services and quality care while admitted. High-risk conditions including Chronic Obstructive Pulmonary Disorder (COPD), Heart Failure (HF), Acute Myocardial Infarction (AMI), Pneumonia, Coronary Artery Bypass Graft (CABG), and Elective Primary Total Knee Arthroplasty/Total Hip Arthroplasty (TKA/THA).

CMS概述了被认为有再入院高风险的健康状况，要求入院时改善服务和优质护理的协调。 高危情况包括慢性阻塞性肺疾病(COPD)，心力衰竭(HF)，急性心肌梗塞(AMI)，肺炎，冠状动脉搭桥术(CABG)和择期原发性全膝关节置换/全髋置换(TKA / THA )。

Risk models were eventually created to predict risk for hospital readmissions. The accuracy of predicting readmissions varies across predictive modeling techniques. This literature review will discuss the most common predictive modeling strategies and the accuracy of predicting hospital readmissions.

最终创建了风险模型以预测医院再次入院的风险。 预测再入院的准确性因预测建模技术而异。 这篇文献综述将讨论最常见的预测建模策略以及预测住院率的准确性。

Methodology

方法

Conducting research involved the search of several databases as well as direct medical journal search. Databases used to search for relevant articles were: Adelphi University Library Online Database (OneSearch), PubMed, Google Scholar, ScienceDirect, Wiley Online Library as well as British Medical Journal, New England Medical Journal, JAMA Network, Springer Link, and PLOS One. In searching for relevant literature primary keywords used were “Hospital Readmissions”, “Frequent Readmissions”, “Predictive Modeling Hospital Readmissions”, “Predictive Modeling Health Care”, “Chronic Health Conditions”, “machine learning”, “deep neural networks”, “LACE” and “artificial intelligence”. Each article was then reviewed with brief summaries to identify trends/patterns and further assist in the organization of the literature.

进行研究涉及几个数据库的搜索以及医学期刊的直接搜索。 用于搜索相关文章的数据库为：阿德尔菲大学图书馆在线数据库(OneSearch)，PubMed，Google Scholar，ScienceDirect，Wiley在线图书馆以及《英国医学杂志》，《新英格兰医学杂志》，JAMA Network，Springer Link和PLOS One。 在搜索相关文献时，使用的主要关键词是“医院再入院”，“频繁再入院”，“预测性模型医院再入院”，“预测性模型保健”，“慢性健康状况”，“机器学习”，“深度神经网络”， “ LACE”和“人工智能”。 然后，对每篇文章进行简要总结，以识别趋势/模式，并进一步协助组织文献。

Each article was saved and tracked with the assistance of the Zotero Reference Management tool. The search turned up 35 articles matching the keywords, however after reviewing the type of article (ie, relevance, year article was written- articles are written> 20 years were excluded, articles that were not peer-reviewed were also excluded) the articles were then narrowed down to 27. Each article was briefly reviewed with written snapshots of each article. Articles that were selected had a focus on the background of hospital readmissions, the risks/benefits associated with predictive modeling in health care, and the predictive modeling techniques currently used.

每篇文章均在Zotero参考管理工具的帮助下进行了保存和跟踪。 搜索结果找到了35个与关键字匹配的文章，但是在检查了文章类型之后(例如，相关性，撰写年份文章-撰写文章>排除了20年，也排除了未经同行评审的文章)。然后缩小到27。每篇文章都经过简短回顾，并附有每篇文章的书面快照。 选择的文章着重于医院再入院的背景，与卫生保健中的预测模型相关的风险/收益以及当前使用的预测模型技术。

The articles were then organized in an order of timeline, starting off with a background of the Hospital Readmission Reduction Program and its significance (this includes systematic reviews that were previously conducted). The organization then continued with the challenges associated with predictive modeling as well as the frequently used predictive modeling techniques.

然后按照时间表的顺序组织文章，首先是“减少医院再入院计划”的背景及其意义(包括先前进行的系统审查)。 然后，组织继续进行与预测建模以及常用的预测建模技术相关的挑战。

Concluding with the most recent articles addressing the more complicated predictive modeling techniques- Deep Neural Networks and Recurrent Neural Networks, as Deep learning presents as the future of predictive modeling in health care. The table below reflects the sources used as part of the literature review, including a breakdown of the type of study, sample size, and predictive modeling method used. Sources primarily used were cohort studies (Prospective and Retrospective), including systematic reviews previously conducted on the topic of predicting hospital readmissions.

最后总结了有关更复杂的预测建模技术的文章-深度神经网络和递归神经网络，深度学习将其视为医疗保健中预测模型的未来。 下表反映了用作文献综述一部分的来源，包括研究类型，样本量和所使用的预测建模方法的细目分类。 最初使用的来源是队列研究(前瞻性和回顾性研究)，包括先前对预测住院率的主题进行的系统评价。

Table Created by Author

作者创建的表

Table Created by Author

作者创建的表

Table Created by Author

作者创建的表

Literature Review

文献评论

The research was conducted to identify common themes associated with readmissions and frequent hospitalizations. At the center of the research common themes discovered were health conditions deemed as high risk for readmission or frequent hospitalizations, and predictive modeling to reduce readmission rates.

进行研究是为了确定与再入院和经常住院有关的共同主题。 在研究的中心，发现的共同主题是被视为再入院或经常住院的高风险的健康状况，以及旨在降低再入院率的预测模型。

The 27 articles were reviewed, the majority of which addressed health condition Heart Failure as the cause for readmissions. The readmission focus time frame was readmissions within 30 days from the discharge date. In an effort to reduce readmissions, predictive modeling techniques have taken the forefront in the health care industry.

对27篇文章进行了审查，其中大多数将健康状况心力衰竭作为再次入院的原因。 重新入院的重点时间是从出院之日起30天内的重新入院。 为了减少再入院率，预测建模技术已在医疗保健行业中走在前列。

The health care industry has sought various methods to predict hospital readmissions in an effort to comply with the Centers for Medicare and Medicaid Services (CMS) Hospital Readmission Reduction Program (HRRP). Healthcare institutions that fail to comply with standards set forth by CMS- frequent “preventable” hospital readmissions will experience penalties and reduced Medicare reimbursements. Withheld payments of up to 3%. Through the research conducted several predictive modeling methods were discovered that were commonly used for predicting hospital readmissions (LACE, Logistic Regression, Support Vector Machine, Cox Proportional Model, Random Forest, eXtreme Gradient Boost, and Deep Neural Networks). Other models such as Naïve Bayes Algorithm, Swarm Intelligence, Survival Analysis, Dynamic Random Survival Forests, and HOSPITAL were also noted, but excluded as part of the literature review.

为了符合医疗保险和医疗补助中心(CMS)的医院减少入院率计划(HRRP)，医疗保健行业已寻求各种方法来预测医院的入院率。 不符合CMS规定的医疗机构-频繁的“可预防”住院再住院将受到处罚并减少Medicare的报销。 最高预扣款项的3％。 通过这项研究，发现了几种预测模型方法，这些方法通常用于预测医院的入院率(LACE，Logistic回归，支持向量机，Cox比例模型，随机森林，极限梯度增强和深度神经网络)。 还提到了其他模型，例如朴素贝叶斯算法，群体智能，生存分析，动态随机生存森林和医院，但未纳入文献综述。

预测建模方法 (Predictive Modeling Methods)

LACE Predictive Assessment Tool

LACE预测评估工具

LACE is a risk readmission assessment tool commonly used by hospitals and has been considered the standard tool of the health care industry. The tool takes into consideration common factors associated with readmissions, which form the acronym LACE.

LACE是医院常用的风险再入评估工具，被认为是医疗行业的标准工具。 该工具考虑了与再入学相关的常见因素，这些因素构成了缩写LACE。

L= Length of Stay

L =停留时间

A= Acuity of Admission

A =录取率

C= Comorbidities

C =合并症

E= Emergency Department Visits

E =急诊科访问

Upon completion of the assessment tool, scores are calculated to determine the risk of readmission. A score of 10 and greater would determine that the patient has a high risk for readmission.

评估工具完成后，将计算分数以确定再次入院的风险。 10分或更高的分数将确定患者再次入院的风险很高。

Logistic Regression Model

逻辑回归模型

Logistic Regression is a Machine Learning (ML) predictive modeling technique and is one of the more widely used ML techniques. Logistic Regression in comparison to other ML techniques is the least complex, which is the attributing factor to why it is more widely used. Logistic Regression is used to analyze data sets and establishing relationships between variables. Dependent variables with Logistic Regression are binary and are used to answer simple questions ie, yes/no questions.

Logistic回归是一种机器学习(ML)预测建模技术，并且是更广泛使用的ML技术之一。 与其他ML技术相比，Logistic回归最不复杂，这就是为什么它被更广泛地使用的原因。 Logistic回归用于分析数据集并建立变量之间的关系。 Logistic回归的因变量是二进制的，用于回答简单问题，即是/否问题。

Support Vector Machine

支持向量机

Support Vector Machine also referred to as SVM, is an ML predictive modeling technique, in which the algorithm focuses on locating decision boundaries and the classified data points. The data points are the support vectors. Decision boundaries also known as Hyperplanes can take shape in the form of a line or a 3-dimensional plane. The distance of the data points from the hyperplane can either be small or large. The size of the distance determines the level of generalization error.

支持向量机也称为SVM，是一种ML预测建模技术，其中算法着重于定位决策边界和分类数据点。 数据点是支持向量。 决策边界也称为超平面，可以以直线或3维平面的形式成形。 数据点与超平面的距离可以小也可以大。 距离的大小决定了泛化误差的程度。

Cox Proportional Model

考克斯比例模型

Cox Proportional Model also is known as the Cox Proportional Hazards Model is another frequently used predictive modeling technique. The Cox Proportional model is a Regression Model that is frequently used in health care analysis. The analysis conducted is also referred to as Survival Analysis, used to predict outcomes or probability of a particular event occurring given the information collected. The outcome is referred to as the hazard rate and the factors influencing the outcome is referred to as co-variates.

Cox比例模型也称为Cox比例危害模型，是另一种常用的预测建模技术。 Cox比例模型是在医疗保健分析中经常使用的回归模型。 进行的分析也称为生存分析，用于根据给定的信息来预测特定事件的结果或概率。 结果称为危险率，影响结果的因素称为协变量。

Random Forest

随机森林

Random Forests Model is an ML predictive modeling technique that utilizes clusters of decision trees to classify data. Each individual cluster generates a prediction. The most common generated prediction is used as the overall prediction for the model.

随机森林模型是一种ML预测建模技术，它利用决策树的群集对数据进行分类。 每个单独的群集都会生成一个预测。 最常见的生成预测用作模型的整体预测。

eXtreme Gradient Boost

极限梯度提升

eXtreme Gradient Boost also known as XGBoost, is another ML predictive modeling technique that utilizes decision trees and Gradient Boosting. Gradient Boosting involves smaller and “weak” decision trees in comparison to Random Forest. XGBoost expands on this modeling technique by “pruning” the decision trees, removing irrelevant information.

eXtreme Gradient Boost也称为XGBoost，是另一种利用决策树和Gradient Boosting的ML预测建模技术。 与“随机森林”相比，梯度提升涉及较小和“弱”的决策树。 XGBoost通过“修剪”决策树，删除不相关的信息来扩展这种建模技术。

Deep Neural Network

深度神经网络

Deep Neural Networks (DNN) also known as an Artificial Neural Network (ANN) is one of the more complex ML predictive modeling techniques. Mimicking the biological structure of the human brain, DNN differs from traditional ML methods as the algorithms used are structured in layers forming essentially forming a “brain” or artificial neural network (ANN). With an ANN the algorithm structure has the ability to learn from the data and make independent decisions and predictions.

深度神经网络(DNN)也称为人工神经网络(ANN)，是较复杂的ML预测建模技术之一。 DNN模仿了人类大脑的生物结构，与传统的ML方法不同，因为所使用的算法被构造为层，这些层实质上形成了“大脑”或人工神经网络(ANN)。 借助人工神经网络，算法结构可以从数据中学习并做出独立的决策和预测。

缺点和局限性 (Drawbacks and Limitations)

Each of the above mentioned predictive modeling techniques have their benefits and drawbacks. The LACE index model as previously mentioned in this paper is widely used across many institutions to determine the risk of unplanned hospital readmission.

上述每种预测建模技术都有其优点和缺点。 如前所述，LACE指数模型已在许多机构中广泛用于确定计划外医院再次入院的风险。

LACE is simple and easy to understand in comparison to ML techniques. One of the key takeaways of LACE is that it takes into consideration co-morbidities, rather than just focusing on one chronic health condition. Research suggests that LACE may perform better at predicting Emergency Room visits rather than admissions.

与ML技术相比，LACE简单易懂。 LACE的主要优点之一是它考虑了合并症，而不仅仅是关注一种慢性健康状况。 研究表明，LACE在预测急诊室就诊率方面可能比入院率更好。

A high LACE score of > 10 did not conclude with hospital readmission, although the risk of readmission is identified. Area Under Curve (AUC) which determines the accuracy of testing, a test closer to 1 represents accurate results, whereas a test that is .5 or lower is considered an inaccurate test. On average, the studies reviewed LACE AUC scores were significantly lower in comparison to other predictive modeling strategies, results varying from 59% (0.59) to 62.8% (0.628). Models mentioned in this paper- Cox Proportional, Logistic Regression, Support Vector Machine, Random Forests, Deep Neural Networks, and eXtreme Gradient Boost performed with higher accuracy in comparison to the LACE index, however accuracy varied from 70% (0.7) to 84% (0.84). The primary limitation reported was the accessibility of data. EHR technology has only recently become more mainstream, there for access to patient data goes back to about 15 years. Although EHR technology has become more mainstream, each system differs, and data collected is not uniform.

尽管已确定有再次住院的风险，但LACE得分> 10的高分并不意味着再次入院。 曲线下面积(AUC)决定测试的准确性，接近1的测试表示准确的结果，而0.5或更低的测试被认为是不准确的测试。 平均而言，与其他预测性建模策略相比，研究回顾的LACE AUC得分明显更低，结果介于59％(0.59)至62.8％(0.628)之间。 本文中提到的模型-Cox比例，逻辑回归，支持向量机，随机森林，深层神经网络和极限梯度增强与LACE指数相比具有更高的精度，但是精度从70％(0.7)到84％不等(0.84)。 报告的主要限制是数据的可访问性。 EHR技术直到最近才变得更加主流，访问患者数据的时间可以追溯到15年前。 尽管EHR技术已成为主流，但每个系统都不同，并且收集的数据也不统一。

This places a challenge for data collection, as some systems contain more patient information than others. Another limitation that has become apparent in the research conducted for the literature review is the impact/role socio-economic factors play in hospital readmissions. Patient insurance coverage, income, housing status all play a role in patient healthcare outcomes.

这给数据收集带来了挑战，因为某些系统比其他系统包含更多的患者信息。 在进行文献综述的研究中，另一个明显的局限性是医院再入院的影响/角色社会经济因素。 病人的保险范围，收入，住房状况都在病人的医疗保健结果中发挥作用。

结论 (Conclusion)

Reviewing the current literature that was researched for this paper, there is consensus that predicting hospital readmissions continues to be a work in progress. We have not reached the capability of achieving an AUC of over 90% (0.9) or close to 1.

回顾当前针对本文研究的文献，人们一致认为预测医院的再入院率仍在进行中。 我们尚未达到AUC超过90％(0.9)或接近1的能力。

There are a number of contributing factors that have an impact on predicting hospital readmissions, ultimately affecting AUC results. Data collection is the primary factor. The research conducted suggests that social and economic factors play a role in readmissions, but this information is typically not including in data collection.

有许多因素影响预测住院率，最终影响AUC结果。 数据收集是主要因素。 进行的研究表明，社会和经济因素在重新录取中起一定作用，但是这些信息通常不包括在数据收集中。

The source of the data collection, EHR systems vary and the composition of the data collected from one EHR may vary greatly from data collected from another EHR system. As an alternative option where data is more uniform in nature, there has been an increased use of Health Information Exchange cloud Data (HIE) to predict risk for hospital readmission. The use of HIE cloud data provides a greater sampling pool for data collection. HIE’s operate as a network containing institutions, hospitals, clinics, primary care offices sharing data for the purpose of improving care coordination and communication between providers.

数据收集的来源，EHR系统各不相同，从一个EHR收集的数据组成可能与从另一个EHR系统收集的数据大不相同。 作为数据本质上更统一的替代选择，越来越多地使用Health Information Exchange云数据(HIE)来预测医院再次住院的风险。 HIE云数据的使用为数据收集提供了更大的采样池。 HIE是一个由机构，医院，诊所，初级保健办公室组成的网络，可以共享数据，以改善医疗机构之间的医疗协调和沟通。

The various ML predictive modeling techniques mentioned in the articles collected, DNN is suggested to present the most accurate predictions for hospital readmissions. Accuracy however falls short of 90% (0.9). The percentage of accuracy AUC for DNN varies from 70% (0.7) to 84% (0.84).

收集的文章中提到了各种ML预测建模技术，建议DNN为医院再入院提供最准确的预测。 但是，准确度不足90％(0.9)。 DNN的准确AUC百分比从70％(0.7)到84％(0.84)不等。

Further research is needed for examples that yield higher results. Logistic Regression continues to be the most commonly used Machine Learning predictive modeling technique, as it is the least complex and easy to understand. However, healthcare as it continues to become more apparent is complex as there is no one size fits all. Research also suggests that predicting hospital readmissions using machine learning techniques generate higher rates of prediction accuracy for single chronic health conditions compared to patients that have multiple comorbid conditions.

对于产生更高结果的示例，还需要进一步研究。 Logistic回归仍然是最简单的机器学习和易于理解的方法，仍然是最常用的机器学习预测建模技术。 但是，由于没有一种尺寸能适应所有情况，因此随着越来越明显的医疗保健变得复杂。 研究还表明，与具有多种合并症的患者相比，使用机器学习技术预测医院的再入院率对单个慢性健康状况的预测准确性更高。

Overall, further research is needed to determine if predicting hospital readmissions with over 90% (0.9) accuracy is feasible. As previously mentioned an array of factors must be taken into consideration in regards to hospital readmissions. The increased presence of EHR technology as well as the use of HIE’s are essential for identifying patient risk for hospital readmission.

总的来说，需要进一步的研究来确定以90％(0.9)以上的准确性预测医院再入院是否可行。 如前所述，在医院再入院方面必须考虑一系列因素。 EHR技术的日益普及以及HIE的使用对于确定患者再次住院的风险至关重要。

Machine Learning presents as a promising option, however, there continues to be room for improvement. The future outlook of predictive modeling involves consistent access to patient data, ideally data that is uniform and does not vary from system to system. The future outlook also involves taking into consideration the patient as a whole- including socio-economic factors as well as comorbid conditions (mental health, substance abuse, etc) and generating a risk score based upon the information collected.

机器学习是一个有前途的选择，但是，仍然存在改进的空间。 预测建模的未来前景涉及对患者数据的一致访问，理想情况下是一致的数据，并且各个系统之间的数据不变。 未来的展望还包括考虑患者的整体情况，包括社会经济因素以及合并症(精神健康，药物滥用等)，并根据收集的信息得出风险评分。

翻译自: https://medium.com/towards-artificial-intelligence/use-of-predictive-modeling-techniques-to-predict-hospital-readmissions-7b12e9f6194e

点击率预测问题建模