Machine Learning and Artificial Intelligence for Personalized Medicine

We consider the estimation of average treatment effects in observational studies without the standard assumption of unconfoundedness . We propose a new framework of robust causal inference under the general observational study setting with the possible existence of unobserved confounders. Our approach is based on the method of distributionally robust optimization and proceeds in two steps. We first specify the maximal degree to which the distribution of unobserved potential outcomes may deviate from that of observed outcomes. We then derive sharp bounds on the average treatment effects under this assumption. Our framework encompasses the popular marginal sensitivity model as a special case and can be extended to the difference-in-difference and regression discontinuity designs as well as instrumental variables. Through simulation and an empirical study on diabetic patients, we demonstrate the applicability of the proposed methodology to real-world settings.

Personalized intervention management in healthcare has received a rapidly growing interest in the big-data era yet still is a burgeoning field. A key challenge for personalization in healthcare is data scarcity. The small sample issue makes standard learning methods hard to learn the right policy and/or suffer from large variances. In this research, we develop a novel data-pooling algorithm in the RL context and establish theoretical performance guarantee. We demonstrate its empirical success on a real hospital dataset with an application to reduce hospital readmission rates. In particular, we show that our algorithm alleviates privacy concerns about sharing health data by requiring sharing aggregate statistics only, which (i) opens the door for individual organizations to levering public datasets or published studies to better manage their own patients; and (ii) provides the basis for public policy makers to encourage organizations to share aggregate data to improve population health outcomes for the broader community. This is a joint work with Xinyun Chen and Shanwen Pu. 

The low abundance of circulating tumour DNA (ctDNA) in plasma samples makes the analysis of ctDNA biomarkers for the detection or monitoring of early-stage cancers (e.g., lung cancer) challenging. We show that deep methylation sequencing aided by a machine-learning classifier of methylation patterns enables the detection of tumour-derived signals at dilution factors as low as 1 in 10000. The low amounts of ctDNA permitted by machine-learning-aided deep methylation sequencing could provide advantages in cancer screening and the assessment of treatment efficacy. Based on the work “Ultrasensitive detection of circulating tumour DNA via deep methylation sequencing aided by machine learning”, Nature BME, 2021.

Personalized medicine requires accurate patient-level prediction.  But often what is desired is not merely prediction but counterfactual accuracy: will this change in treatment change the patient’s outcome?  Deep neural networks are proving to be accurate for prediction but not yet for answering causal or counterfactual questions.  This talk presents motivation, theory, and initial methods toward making them so.