Using Machine Learning for Personalized Medicine

医学界越来越关注“个性化医疗”：根据史密斯女士的所有特征，确定最适合她的具体治疗方法。在过去的十年里，基本上我所有的研究都集中在这个挑战上。我和我的实验室使用了“机器学习”方法：根据历史患者的数据集找到相关的分类器。虽然标准的机器学习系统足以学习其中一些任务的分类器，但许多其他任务需要创新的扩展，* (a)在医学图像中找到有用的模式（例如，扩展条件随机字段，使它们能够有效地定位患者MRI扫描中的脑肿瘤），或* (b)处理当今“组学”数据集中的高维数据(例如，基因表达微阵列中的50,000个特征，或患者单核苷酸多态性[SNP]谱中的100万个特征)，或* (c)以适应时间序列数据，包括fMRI数据中的时间谱和帮助患者管理其I型糖尿病的顺序控制成分。*我们还设计了新的基础技术——例如，有效地收集产生有效分类器所需的信息（“预算学习”），以及开发一种学习患者特定生存分布（PSSP；想想史密斯女士自己特有的Kaplan-Meier曲线）的工具。在接下来的五年里，我计划继续在这个令人兴奋的领域工作，通过开发必要的工具来实现个性化医疗。和以前一样，我预计一些任务将涉及到应用标准技术；这些结果仍将在医学上有用，并将有助于向医学界展示机器学习，作为一种产生真正个性化治疗的方法。其他任务将需要进一步扩展先前开发的创新，例如添加PSSP系统（例如，使用l1正则化，以确定每次相关的特征），以及应用强化学习技术来管理糖尿病和其他慢性疾病的其他新方法。许多疾病，如大多数癌症，都具有极强的异质性——不仅存在于人群中，而且存在于单个个体中。我们将探索有效地对这些疾病进行建模的方法，因为这些模型对于确定适合个体的治疗方法至关重要。考虑到收集样本的挑战（和成本），许多训练数据集实际上是来自多个站点的数据的组合。这导致了几个问题：批效应（由于不同实验室之间的技术差异等）和协变量转移（每个子研究中的人口略有不同）。我将继续探索发现并纠正这些微妙但极其重要的问题的方法。**然而，我计划将我的大部分努力用于寻找学习算法的有效方法，以使用先前的生物学知识（例如，代谢组学和/或信号传导途径）来提高学习分类器的性能。考虑到今天的路径信息固有的不完善，这将需要（同时）确定改进路径信息的方法。我的经验表明，这些信息对于处理今天（甚至更多，对于明天）的高维数据（微阵列、snp、CNVs、下一代测序等）至关重要，否则就无法识别与许多多病灶疾病（如癌症）相关的复杂模式。因此，这种分析将需要为应对这些重要疾病的患者提供个性化治疗的工具。

The field of medicine is increasingly focusing on "personalized medicine": identifying the specific treatment that is best for Ms Smith, based on all of her attributes. Over the last decade, essentially all of my research has focused on this challenge. My lab and I have used the "machine learning" approach: finding relevant classifiers, based on datasets of historical patients. While standard machine learning systems were sufficient to learn classifiers for some of these tasks, many others require innovative extensions,* (a) to find useful patterns in medical images (eg, extending conditional random fields to enable them to effectively locate brain tumors within a patient's MRI scan), or* (b) to deal with the high dimensional data in today's "omics" datasets (eg, the 50,000 features in a gene expression microarray, or the 1 million features in a patient's Single Nucleotide Polymorphism [SNP] profile), or* (c) to accommodate time sequence data, both the time spectra in fMRI data, and the sequential control components in helping a patient manage her Type I diabetes.*We have also designed new foundational technologies -- eg, to efficiently collect the information required to produce effective classifiers ("budgeted learning"), and to develop a tool that learns a patient specific survival distribution (PSSP; think of a Kaplan-Meier curve, specific to Ms Smith herself).**Over the next five years, I plan to continue working in this exciting area of enabling personalized medicine by developing the necessary tools. As before, I anticipate some tasks will involve applying standard techniques; these results will still be medically useful, and will help demonstrate machine learning to the medical community, as a way to produce truly personalized treatments. Others tasks will require further extensions to the innovations developed earlier -- such as additions to the PSSP system (eg, using L1-regularization, to identify which features are relevant at each time), and additional novel ways to apply reinforcement learning techniques to manage diabetes, and other chronic diseases. Many diseases, such as most cancers, are extremely heterogeneous -- not just wrt a population of subjects, but also within a single subject. We will explore ways to effectively model such diseases, as these models will be essential to identify the appropriate treatment for the individuals. Given the challenges (and costs) of collecting samples, many of the training datasets will actually be combinations of data from multiple sites. This leads to several problems: batch effects (due to technical differences between the different labs,etc.), and covariate shift (slightly different populations in each sub-study). I will continue to explore ways to detect, and then correct, these subtle, but extremely important, problems.**I plan to devote most of my efforts, however, to finding effective ways for learning algorithms to use prior biological knowledge (eg, of metabolomic and/or signaling pathways) to improve the performance of learned classifiers. Given the imperfections inherent in today's pathway information, this will require (in parallel) identifying ways to improve that pathway information. My experience has shown that such information is essential to deal with today's (and even more-so, for tomorrow's) high-dimensional data (of microarrays, SNPs, CNVs, next generation sequencing, etc.), as there will, otherwise, be no way to identify the complex patterns associated with many multi-focal diseases, such as cancer. Hence, such analyses will be required to produce tools for personalized treatment for patients coping with these important diseases.