October 16, 2019
Eric Rose, PhD
A goal of precision medicine is to tailor treatment to the unique characteristics of a patient. For example, when there are multiple treatment options or strategies available, how does one select the “best” treatment for an individual? Ideally, that decision is based on the available patient information. But understanding the relationship between a patient’s unique characteristics and the effectiveness of a treatment is a complex problem. This is particularly true for chronic diseases such as cancer where treatments are assigned sequentially over a period of time. In these situations, we are commonly interested in maximizing the survival time of a patient. This causes complications for analysis though since when we are studying a time-to-event, such as survival time, we will have patients that we do not observe the event for. This is called censoring, which can cause naïve application of standard statistical methods to produce biased results.
Frequently in biostatistics, we are interested in comparing different potential treatments to try to determine which would be best for future patients. To do this, we must have some formal outcome or set of outcomes of interest that we use to define what treatment is best. This outcome changes based on the context of what we are studying. For example, there may be some continuous measure of severity that we want to minimize or we may be interested in the presence/absence of a disease after some amount of time.
Another possible outcome of interest is the amount of time a patient survives with a disease. This is an example of a time-to-event outcome. Other possible time-to-event outcomes are the time until a patient is in remission or the time to the reccurrence of a disease. These types of outcomes are common when studying chronic diseases such as cancer or cardiovascular disease.
When we are dealing with time-to-event outcomes we frequently have patients in our study for whom we never observe the event we are interested in. This could be because the patient drops out of the study midway through or the study is of a limited length so the patient hasn’t had the event occur by its completion. We call an observation censored if we do not observe when the patient experienced the event. If we ignore the patients that are censored when conducting our analyses, we could potentially end up with biased results. The area of statistics that covers how to deal with time-to-event data is called survival analysis.
Now suppose we are assigning multiple treatments to a patient sequentially over time and are interested in tailoring treatments for patients based on individual patient characteristics. This is called a dynamic treatment regime, which is defined as a set of rules, one for each decision point, that maps patient characteristics to a recommended treatment. Our goal in this context is to estimate the optimal treatment regime that would maximize our outcome of interest if all patients in the population of interest were assigned treatments via these rules.
In this context our outcome of interest could also be a time-to-event in which case our usual methods would still lead to bias as before. In this case though there are further considerations that need to be accounted for. This is because each patient will not necessarily reach all of the different decision points. This could be because either the patient was censored during the course of the study or they had the event occur before reaching all decision points.
These problems create a difficult challenge for statisticians to study. Overcoming some of these obstacles has been an active area of research in recent years. With time-to-event outcomes being so prevalent in certain areas of medicine, furthering our understanding of how to estimate optimal dynamic treatment regimes in this scenario has great potential to improve our understanding of how to best treat patients to improve health care.
Eric has just defended his PhD thesis (congratulations, Dr. Eric Rose)!! We asked a fellow Laber Labs colleague to ask Eric a probing question.
It is said that indifference, not hate, is the opposite of love. What is the opposite of pancakes?
It seems obvious to choose the natural nemesis of the pancake, the waffle. For centuries, these two foods have competed for those looking for a sweet, doughy treat. But just as love and hate on the surface appear to the antithesis of each other, they are in fact very similar. The natural transition is then to switch to one of the many savory foods eaten for breakfast around the country every day. I believe though that the ideal place for a pancake is as a sweet side to complement a plate of bacon and eggs. To me the true opposite would be something that fills the same role within the meal, but is far different in composition or taste. Therefore, the true opposite of pancakes is plain white toast.