Eric Laber
The vision for personalized medicine is a healthcare system that is deeply tailored to each individual patient leading to better health outcomes while reducing cost and resource consumption. Increased access to large amounts of patient data in the form of health and billing records, insurance claims, and adverse event reporting databases etc., offer an unprecedented opportunity to use data to inform personalized medicine. Thus, statisticians have a key role to play in the transition into the new era of personalized medicine.
In the statistics literature, a central concept in the study of personalized medicine is that of a treatment regime. A treatment regime is a mathematical model of the clinical decision process wherein a clinician uses accruing patient information to adapt treatment to a patients changing health status. This model is represented as a sequence of decision rules, one per stage of clinical intervention, that map up-to-date patient information to a recommended treatment. Coupling a treatment regime with a notion of optimality, i.e., maximal expected symptom reduction, admits a formal definition of a `best regime' and subsequently allows the statistical crank to be turned to grind data into an estimated optimal treatment regime. Furthermore, this framework makes it possible to study a large number of variants of the treatment regime estimation problem, e.g., missing or censored observations, measurement error, longitudinal data, shifting covariate distributions, high-dimension, and so on. Statisticians have been working feverishly over the past 15 years to lay the theoretical and methodological foundations of data-driven treatment regimes including the foregoing variations. This work has produced a number of fundamental insights in the nature of data-driven decision making and generated new knowledge in causal inference, machine learning, optimization, and semi-parametric theory. However, it is time for statisticians to think carefully about the direction of methodological research in personalized medicine.
In the quest for (mathematically) optimal data-driven treatment regimes, there has been a push toward more flexible estimators. At the extreme, these estimators become a complex series of computationally intensive steps producing a blackbox treatment regime that is hardly intelligible even to estimator's architect. In many settings, as the estimated regime becomes more complex it becomes less likely it is to impact actual clinical practice. One reason for this is that no prudent clinician is going to take a black-box statistical model fit using a single data set and use it to assign treatments to patients. It is foolish to think that a statistical model fit to data from a single randomized or observational study should supplant both clinical expertise and all past scientific evidence that would otherwise bear on a treatment decision. Yet, in the statistical community, we often motivate the development and evaluation of methodology for treatment regimes as if clinical scientists will blindly abide by the resulting estimated regime. Instead, an estimated regime is more likely to (and should) be used to inform additional research and to become one of many pieces of information contributing to an eventual scientific consensus on how treatment should be tailored to individual patients.
Moving forward, it is imperative that statistical research on treatment regimes recognize the role that data-driven regimes are likely to play in scientific development. If our goal is to impact clinical practice, we must consider: (i) the interpretability of a regime; (ii) the ability to quantify trade-offs across multiple outcomes, i.e,. side-effects, measures of efficacy, cost, burden, etc.; (iii) using data to generate new clinical hypotheses; and (iv) visualization and other means of communicating information to patients and clinicians. P01 investigators working on treatment regimes are beginning to make headway on some of these research directions but there is much work to be done. We encourage and welcome both clinical and statistical scientists to help us push this research forward.
Comments/Discussion
Scientists have been thinking about how to utilize the big amount of data made available by the current technologies and sciences. I think not only innovation on clinical trials, but also innovation on how diagnosis and treatment assignment is performed, will be desired. A science fiction scene: in the near future, patients will get themselves treatments at home using a mobile app on their smart phones.
A. Shan
Submit a question or comment. |
Each month, one of our program investigators will introduce him/herself and will discuss their research, new research directions, or advancements made toward our goal of improving the clinical trial process. Readers are encouraged to send questions or comments. In addition, we will announce new software releases, publications, and upcoming events.
Subscribe |
Date: February 15, 2016
Date: January 15, 2016
Date: December 16, 2015
Date: November 24, 2015
Date: October 15, 2015
Date: September 17, 2015
Date: August 24, 2015
Date: July 16, 2015
Date: June 4, 2015