关键词： 医学数据   算法   糖尿病  

摘要： 目的探索机器学习算法及衍生算法在医学数据集上的分类效果,以期更好的发现计算机在辅助医学诊断方面的应用价值。方法以皮马印第安人糖尿病数据集为例,利用WEKA平台构建机器学习模型,包括基于贝叶斯定理的NavieBayes、基于集成学习的Bagging、基于树思想的J48等模型,共六大类21种算法,运用多维度多指标对所建立模型的预测效果进行评价。结果RMSE和RRSE均较小的前5位算法依次为Logistic、LMT、RotationForest、RandomForest和Bagging;LMT、SMO、Logistic、NavieBayes、RotationForest的分类正确率均超过了76%,其真阳性率均在76%以上,ROC曲线显示,除SMO外,其余算法曲线下面积均在0.82以上。结论在该糖尿病数据集上的分类预测效果较好的算法有6种,分别是LMT、SMO、Logistic、NavieBayes、RotationForest和Bagging,均具有较高的正确率和预测价值。

摘要： Recent advances in machine learning (ML) hold much promise for using data to drive more effective decisions. However, many challenges remain to our realizing this decision-making potential, given the limitations of predictive algorithms and imperfections in the data available. This thesis investigates these critical challenges in the areas of data-driven optimization, causal inference, and algorithmic fairness, and develops fundamental theory and new ML methods. In Part I, we focus on data-driven optimization that involves both uncertain quantities of interest (e.g., demand) and predictive contextual features (e.g., product characteristics). Chapters 2 and 3 focus on stochastic optimization problems and investigate two popular paradigms: an "estimate-then-optimize" paradigm that first uses standard ML methods to predict the distribution of uncertain quantities given contextual features and then plugs the distributional predictions into a stochastic optimization problem to solve for decisions, and an "end-to-end" paradigm that integrates the prediction and the decision-making aspects by directly training predictive models to target good decisions. In Chapter 2, we develop an "end-to-end" stochastic optimization forest algorithm that constructs decision trees to directly optimize decision quality, which we show provides significant benefits for decision making over building trees that target prediction. In Chapter 3, we show a more nuanced landscape for the integration of estimation and optimization, by identifying many common settings where "end-to-end" approaches can actually have much slower regret-convergence rates than the far simpler "estimate-then-optimize" approach. Chapter 4 considers the online setting where we collect more data as we make decisions and studies nonparametric contextual bandits with smooth expected reward functions. We develop a novel algorithm that leverages this smoothness structure and show that its regret rate is minimax optimal. Our regr

关键词： 卵巢癌   计算机辅助诊断   机器学习   卷积神经网络   核磁共振成像  

摘要： 卵巢癌(Ovarian Cancer)是女性生殖系统常见恶性肿瘤,其致死率居妇科恶性肿瘤首位。其中,上皮性卵巢癌(Epithelial Ovarian Cancer,EOC)为主要类型(占比>90%)。术前准确评估EOC状态,可帮助患者实施个性化精准治疗,提高其生存率。相比于超声和CT等常用成像技术,核磁共振成像(Magnetic Resonance Imaging,MRI)在EOC分期及复发转移评估等方面具有明显优势,是EOC首选的影像学术前诊断方法。然而,EOC影像表现异常复杂,基于肉眼的临床诊断主观性高,极易误诊。机器学习方法可借助大量样本挖掘图像中的深层次特征,构建模型提高诊断的准确率,有效减少误诊和漏诊。本文利用机器学习方法,针对EOC分割、诊断中的一系列问题开展大规模多中心研究,主要工作如下:(1)针对EOC病灶人工标注工作量大、难度高的问题,提出一种用于EOC自动分割的多边输出全卷积神经网络。该网络使用多个边输出模块对提取到的不同层级特征进行解码,以生成包含不同尺度信息的边输出分割结果,最终将其融合并生成更为精细的最终分割结果。所提模型能够减少临床医生的工作量,提高病灶勾画的准确性和稳定性。(2)针对Ⅰ型和Ⅱ型EOC临床鉴别困难的问题,提出一种基于多模态MRI的影像组学-机器学习模型。从患者的多模态MRI(T2WIFS,DWI,ADC和CE-T1WI)数据中提取形状特征、小波特征等八类特征,并分别使用支持向量机、决策树和线性模型构建多模态的混合模型。此外,我们还使用遮挡测试对模型的重点关注区域进行识别,该方法有望用于辅助术中冰冻病理切片采样的定位。(3)提出一种用于EOC和BEOT鉴别的新型多示例卷积神经网络。所提模型能够借鉴临床医生的决策方式,自动感知不同模态对决策的重要程度,并基于其重要性实现多模态MRI特征融合。此外,该模型还将肿瘤分布的强先验知识作为决策的重要参考,在对示例进行预测时参考其邻近示例的预测值,从而实现更加精准的预测。

摘要： Precision medicine promises the ability to intelligently tailor clinical interventions to individual patient needs. Personalizing clinical management will necessitate access to high quality, accurate, and functional information about disease state. However, traditional noninvasive diagnostics such as imaging rely on static measurements of tissue morphology, while molecular features of disease can often only be discerned through invasive methods like biopsy. To acquire functional, personalized information about a disease, a promising approach would be to engineer sensing probes that can directly query disease activity to generate amplified signals that can be analyzed externally. To this end, we have engineered responsive nanosensors that probe tissue microenvironments and detect dysregulated protease activity as a functional biomarker of disease. Because proteases directly contribute to multiple disease processes, including cancer, these sensing probes have the potential to enable quantitative, dynamic, and personalized monitoring of disease activity. In this thesis, we designed new classes of these protease activity sensors and machine learning methods to accurately detect cancer and to functionally profile its biology, aiming to promote the utility of these tools for precision medicine. We first established diagnostic platforms that leverage responsive nanoparticles to measure protease activity in vivo and release reporters that can be detected in the urine. We exploited this basic framework to invent a simple, fast color change urinary readout, and further demonstrated its translational potential by integrating nanosensor multiplexing and machine learning to accurately detect localized lung cancer in two genetically engineered mouse models of lung adenocarcinoma. We then sought to establish generalizable pipelines to rationally design such tools and to functionally dissect the biology of protease dysregulation in cancer. We designed an integrated suite of ex vivo

摘要： Snowfall forecasting has historically been an area of difficulty for operational meteorologists, particularly in the remote complex terrain of the Western US. Attempts at improving forecasts have been made, but skill is still poor, with snowfall routinely overpredicted. A major reason for this overprediction has been a failure to accurately predict snow-liquid ratios (SLR) ahead of major events. This research proposes, develops, and tests multiple machine learning methods for dynamic SLR prediction for the Sky Islands of southeast Arizona in the form of a simple feed-forward neural network, a support vector machine, and a k-nearest neighbor algorithm. Input parameters were chosen based on variables found by previous studies to have a regression-based relationship with SLR, with a focus on the lower-mid levels of the troposphere. These parameters were also used to construct a multiple linear regression model, and its performance was compared with the machine learning methods. Each of the machine learning methods showed significant improvement compared to the multiple linear regression. When tested on historical events, nearly 95% of the network-predicted SLR values fell within the margin of error of observed SLRs, calculated using verification data from Zeng et al (2018), with slightly higher accuracies for both the SVM and KNN algorithms. Each showed significant gain in skill compared to the multiple linear regression model. Current and future work is focusing on shifting to higher resolution data, as well as adjusting the model to look at a wider region within the mountainous Western US in order to achieve greater operational benefit.

摘要： When making a decision about whether a particular treatment is beneficial for a patient, the only published information available is based on the average effects across many subjects reported in a clinical trial research. Doctors and patients alike would like to move toward a promising future of treatment in which decisions can be improved via patient's personalized information such as demographics, medical history, and lab test results. However, this is a challenging causal inference problem, known as individual treatment effect (ITE) estimation. While many methods have been recently developed to predict individual treatment effects from clinical trial data, a standardized and comprehensive evaluation framework is urgent *** this study, we develop a new pipeline for training and evaluating ITE models with two useful new features. First, we apply nested cross validation to allow flexible and effective hyperparameter tuning while producing non-contaminated estimates of heldout model performance even on small trials datasets. Second, we use multiple performance metrics to assess both a model's ability to predict outcome and treatment effect accurately and the calibration of their *** demonstrate the application of our new pipeline in a case study of results from a randomized trial(Diabetes prevention program) in which lifestyle interventions and metformin treatment were considered to avoid the onset of diabetes. In this case study, we find that when properly compared, more flexible prediction models designed specifically for ITE based on causal boosting and causal random forests did not deliver any better ITE predictions than generalized linear models. We hope our work inspires future improved results and ultimately moves the world a little closer to the promise of personalized medicine.

关键词： Logistics   Data analytics   Uncertainty   Machine learning   Miles per stop prediction  

摘要： Logistic providers have learned to efficiently serve their existing customer bases with optimized routes and transportation resource allocation. The problem arises when there is potential for logistics growth in an emerging market with no previous data. The purpose of this work is to use industry data for previously known and well-documented markets to apply data analytic techniques such as machine learning to investigate the uncertainty in a new market. The thesis looks into machine learning techniques to predict miles per stop given historical data. It mainly focuses on Random Forest Regression Analysis, but concludes that additional techniques, such as Polynomial Regression are promising for this problem. Additionally, data processing and cleansing is implemented on a model different than what is currently being used by the logistic provider. The results indicate that through the use of polynomial regression on pre-processed and cleaned data, a 75% improvement in performance can be achieve in comparison to the baseline established by the logistics provider.

关键词： 机器学习   共享住宿   价格预测   Gradient Boosting模型  

摘要： 机器学习涉及多个领域,涵盖了多门学科知识,属于新兴的交叉学科,统计学、概率论等学科构成了其基础架构,其本质就是在此基础之上,在计算机当中运用这些知识来模拟或者进一步实现人类的学习行为。机器学习在近些年随着互联网技术的迅猛发展而广为人知,但其发展历史并不是如想象中那般短暂,其实早在几十年前机器学习就已经存在,甚至可以追溯到几个世纪之前,机器学习的基础架构就已然得到了构筑。像17世纪产生的最小二乘法的推导,以及马尔可夫链等,这些知识的发现使得机器学习能够在今天得到广泛的应用与研究。而在当今的时代背景下,大数据、人工智能等领域飞速发展,使得机器学习的应用前景也愈加广泛,因此本文即以机器学习为手段,来进行短租房价格预测方面的研究。随着经济社会的不断发展,不断进步,越来越多的经济活动以更新颖的方式走进大众生活,社会也需要更加有效的方式来进行资源配置,共享经济应运而生。从共享经济提出一直发展至今,共享一词几乎涉及了社会生活的方方面面,例如出行、住宿、办公、学习等,而共享住宿随着近些年由国外流入国内,逐渐得到发展,也在中国大地上刮起了一股在线短租的风潮,极大地方便了人们的生活、工作。并且由于其个性化的服务,价格的优势,以及更具人性化的特点,日益成为人们出行的首要选择,中国的共享住宿市场也得到了空前的繁荣与发展。本文以共享住宿房屋价格为研究对象,通过获取阿里天池竞赛当中的共享住宿短租数据,以Airbnb平台在北京市的房源信息为例,研究各因素与价格的相关关系,明确各影响因素的影响程度大小,确定要放入预测模型当中的变量。在本文当中,具体的研究工作主要有三部分,如下:第一部分:共享住宿数据的预处理工作。在这一部分当中本文主要是对于所获得数据进行各项预处理,包括数据清洗(异常值、空值的处理),冗余属性的处理、共享住宿数据当中定性数据的数据变换(独热编码、标签编码、标准化与平滑处理),从而为后面的建模工作提供可以直接使用的数据。第二部分:共享住宿数据的描述性分析。在这一部分当中,本文对于所获得数据进行描述性分析,将所有的属性列分为三个部分,第一部分为本文的研究对象,共享住宿房屋的价格;第二部分是共享住宿房屋的身份识别属性列;第三部分是可能对共享住宿价格有影响的影响因素属性列。本文在这一部分当中所做的主要工作,就是分析第三部分各属性列与研究对象共享住宿房屋价格的相关程度大小,从而确定最终放入模型当中的特征变量。第三部分:共享住宿价格预测的算法模型。这一部分是本文的重点研究部分,在这一部分当中,本文一共使用了四种机器学习算法模型,分别为Gradient Boosting(梯度提升)、XGBoost、Random Forest(随机森林)、Bagging 回归,并对四种模型用模型得分以及均方误差进行评价,最终发现表现最好的是Gradient Boosting模型,预测结果也能够达到研究的预期。通过上述研究,本文期待能够找出表现较好的算法模型,来对共享住宿房屋的价格进行预测,这种预测工作将会对消费者的消费决定提供支持,也会对房东的定价工作提供参考,具有十分重要的现实意义。本文创新使用了机器学习算法模型来对房屋价格进行预测,脱离了传统价格预测的线性回归模型,使之具有更好的预测准确度。同时在本文当中,也进一步对共享住宿房屋价格的影响因素做了探讨,通过各种可视化工具,使结果能够更加清晰的展现在读者眼前。

关键词： SMOTE过采样   逻辑回归   XGBoost   LightGBM   模型融合  

摘要： 随着我国经济的快速发展与居民超前消费理念的产生,借贷成为企业和个人解决经济问题的重要方式。为了鼓励居民消费,国家大力支持发展信贷产业及其他金融服务。随着金融市场的竞争日益激烈,信贷企业等各类金融机构为了增加收益,就会开发出多种类型的金融产品。如果这些金融企业的风险预警未能及时监控落实,就会给其带来不可估量的经济损失。在这种情况下,建立一个科学有效的风险评估模型对信贷业的健康发展显得尤为重要。本文通过构建集成分类模型,帮助信贷平台对客户进行风险评估,识别违约客户,以最大限度地对有可能遭遇的信用风险进行规避。本文主要通过机器学习分类算法对贷款违约客户进行识别分类,首先对原始数据进行数据处理,包括缺失值和异常值的处理及特征预处理。为了降低模型复杂度和提高模型效率,本文引入了递归特征消除法对特征进行筛选。针对数据不均衡的问题,本文采用了SMOTE过采样方法对传统的分类算法做出了改进,提升对少数违约客户的预测效果。其次结合贪心调参和网格搜索算法对逻辑回归模型、随机森林模型、XGBoost模型和LightGBM模型的相关参数进行修正,提高了模型的泛化能力。最后对分类表现较好的模型进行融合,融合结果通过Stacking方式输出,并在不平衡的测试集上进行模型实验,实验结果表明Stacking集成模型对违约客户的识别效果更好。

关键词： 蛋白质-配体相互作用   蛋白质-蛋白质相互作用   机器学习方法   错义突变  

摘要： 蛋白质与配体的相互作用对生物体内发生的许多过程有着至关重要的作用,触发多种信号转导过程。改变配体与蛋白质结合的错义突变是导致遗传疾病和耐药性的关键机制。此外,药物靶点上的突变也是产生耐药性的主要原因。目前,Rosetta以及炼金术自由能计算方法被广泛用于预测突变引起的蛋白质-配体结合亲和力的变化,但是效果仍不令人满意,需要在准确性和速度上进一步提高。因此,我们开发了在计算成本和精度之间进行权衡的机器学习方法PremPLI,用于定量估计突变引起的亲和力变化。该方法在来自360个复合物的796个突变上进行训练,由11个对模型有显著贡献的特征构建随机森林回归函数。实验测量值与预测值之间的皮尔森相关系数为0.70,对应的均方根误差为1.08 kcal mol-1。我们在两个基准数据集上对PremPLI与其他可用方法的预测性能进行全面比较,证实了我们的方法运行稳定。PremPLI可以相对准确地估计结合亲和力变化并且区分耐药性突变和易感突变,但所需的计算资源却很少。PremPLI 可从 https://***/research/PremPLI/免费获取。蛋白质-蛋白质相互作用在大多数生物学过程中都是必不可少的。目前有很多定量预测突变对蛋白质-蛋白质结合亲和力影响的计算方法,但它们在识别打破蛋白质-蛋白质相互作用的突变方面精确度有限。这里,我们引入一种新的预测方法PremDPPI,用于鉴别打破蛋白质-蛋白质相互作用的突变。PremDPPI使用随机森林分类函数,由10种基于序列和结构的特征组成。PremDPPI模型在三种交叉验证中表现较好,其中准确率不低于0.90,精确度不低于0.86,且马修斯相关系数不低于0.75。PremPLI和PremDPPI有助于找到功能上的重要突变、识别耐药性突变以及发现潜在的药物靶点。