Using Machine Learning for Effective Personalized Treatments

点击翻译按钮获取中文摘要

The field of medicine is increasingly focusing on "evidence-based personalized medicine": identifying the specific treatment that is best for Ms Smith, based on all of her attributes, using a model that is based on earlier patients in a labeled dataset. Over the last 15 years, most of my research has focused on this challenge -- either directly or indirectly. Many projects directly viewed this task as a "machine learning" task of learning an appropriate classifier from a relevant labeled datasets. While standard machine learning systems were sufficient to learn useful models for some of these tasks, many others require innovative extensions -- eg, to cope with data that was extremely high dimensional or sequential in nature, etc. Some other projects produced tools that enable these tasks, by providing relevant features -- eg, techniques that identified and quantified the molecular compounds (“metabolites”) within a biological sample, using mass spec or NMR spectroscopy. Yet others explored foundational topics -- related to predicting individual treatment effect scores, or survival prediction. In addition to many publications, these activities have also led to some deployed medical applications (eg, a commercial tool for screening for adenoma), and several web-apps -- for helping clinicians determine which liver-patient should be added to the waitlist for a new liver, for producing effective survival prediction models (PSSP), and for analysis of NMR spectra (Bayesil) and of ESI-MS/MS and EI MS (CFM-ID). Over the next 5 years, I plan to continue working in this exciting area of enabling evidence-based personalized medicine by developing effective models, and the tools needed for producing them. As before, I anticipate some tasks will involve applying standard techniques; these results will still be medically useful, and will help further convince the medical community that machine learning can be effective, as a way to produce effective personalized treatments. Most of my lab's research efforts, however, will focus on other computationally interesting questions, all towards the goal of producing effective predictive models: (1) Improving approaches for predicting survival time, from datasets with censored instances (2) Learning prognostic models (versus simple diagnostic ones); this will require a better understanding of counterfactual reasoning and other approaches of coping with (biased) training data that involves interventions. (3) Addressing issues from multi-site data, including batch effect and covariate shift (4) Coping with high-dimensional data using various dimensionality reduction techniques, including tools that exploit the structure of the data, such as topic models Each of these topics will require both theoretical analysis as well as empirical experiments. Moreover, while our main application domain is medicine, each of these represents a core machine learning challenge, meaning our results will apply across domains.