SEI: Diagnosis and Treatment of HIV Patients using Data-Mining Techniques: Making Inferences from Imperfect Data

HIV感染的治疗是具有挑战性的，因为快速发展的治疗策略和影响患者的复杂的社会问题。从数据库中提取信息并提供信息以解决治疗策略选择中涉及的一些不确定性的工具是必要的。遗憾的是，现有数据库存在从信息缺失到数据输入错误和主观评估等各种不完善之处。在机器学习中，数据缺陷没有得到足够的重视，这是令人惊讶的，因为几十年来在不确定性处理领域的研究开发了大量的数学框架。在拟议的工作中，一个由三名研究人员组成的团队将通过从相关数据库中收集可操作的知识来解决将增强医生S决策能力的研究任务。这些任务包括检测频繁发生的疾病，这些疾病需要在具有模糊性和不确定性的时变域中进行关联挖掘，预测特定治疗的成功，特别注意从稀疏和不可靠的数据中进行归纳，预测患者对模糊属性的不依从性，以及对知识的统计和医学验证。1)修改现有的关联挖掘技术，使其能够处理歧义，并能够量化结果的不确定性。还将开发反映诱导知识的时变性质的技术。2)开发了一种新的聚类算法(基于协同过滤)，它可以忽略训练样本的描述，并对现有的协同过滤技术进行修改，使其能够处理数据缺陷。3)开发用于从含糊描述的例子中归纳分类器的机器学习技术。这三种贡献都可以用于从不完善的数据库中进行知识发现。在医学领域，归纳出的知识将提供新的假设和新的治疗策略。这项研究项目涉及来自不同学科的专业人员的多学科合作。医科学生将学习欣赏现代计算机科学技术如何提高医疗实践，而工程专业的学生将学习计算机科学在医学应用中遇到的复杂情况。开展推广活动，培养高中生和社区大学生的参与。迈阿密大学是一所西班牙裔服务机构；拟议的研究将涉及工程学研究中代表性不足的学生群体。通过出版物和专题介绍广泛传播研究成果，将理论和实验工作纳入课程，并在相关网站上发表意见。

ABSTRACTTreatment of HIV infection is challenging because of both rapidly evolving therapeutic strategies and the complex social problems affecting patients. Tools that extract information from databases and provide information to resolve some of the involved uncertainties in selection of therapy strategies are necessary. Unfortunately,existing databases are marred with imperfections ranging from missing information to data entry errors and subjective evaluations. In machine learning, data imperfections have received inadequate attention,which is surprising in view of the abundance of mathematical frameworks developed by the decades of research in the .eld of uncertainty processing. In the proposed work, a team of three researchers will address the research tasks that will enhance the physician 's decision-making capabilities by gleaning actionable knowledge from relevant databases. These tasks include detection of frequently co-occurring diseases that require association mining in time-varying domains with ambiguities and uncertainties, prediction of the success of specific treatments, with special attention to induction from sparse and unreliable data, prediction of a patient non-compliance with a focus on ambiguous attributes and statistical and medical validation of the knowledge.The proposed research will contribute to computer science along the following three lines. 1) Modification of existing techniques for association mining so that they can work with ambiguities and can quantify the uncertainty of the results. Techniques that reflect the time-varying nature of the induced knowledge will also be developed. 2) Development of a novel clustering algorithm (based on collaborative filtering) capable of ignoring the descriptions of the training examples, and modification of existing collaborative-filtering techniques so that they can handle data imperfections. 3) Development of machine-learning techniques for classifier induction from ambiguously described examples. All of these three contributions can be used in knowledge discovery from imperfect databases. In the medical domain, the induced knowledge will provide new hypotheses as well as new treatment strategies.This research project involves a multidisciplinary collaboration of professionals fromdifferent disciplines. The medical students involved will learn to appreciate how modern computer science techniques can enhance medical practice,while the engineering students will learn about the complications encountered in medical applications of computer science. Outreach activities to develop the participation of high school and community college students will be developed. The University of Miami is a Hispanic Serving Institution; the proposed research will involve under-represented student groups in engineering research. Activities on broad dissemination of research results via publications and presentations, incorporation of theoretical and experimental work into courses, and by contribution to relevant Internet sites.