Advances in machine learning and statistical theory can now learn from data which interventions best suit which kinds of people. In other words, we now have algorithms that take as input a patients’ characteristics and spit back out the intervention that will yield the best outcome for them. There are two “data” stages needed for carrying out these so-called optimal individualized treatment rule in practice. In the initial data stage, we design the optimal intervention rule: we collect data on 1) patient characteristics, 2) the intervention the patients got in the study, and 3) the outcome each patient had after going through the intervention they got, and we learn the rule or algorithm based on that existing data using data-adaptive (machine learning) methods. In the second data stage, we apply the learned optimal rule on incoming, new patient data to assign an individualized intervention to each person based on their characteristics. In this class, I hope to think about the second stage, i.e., what it would look like to have a machine learning-generated treatment decision inform clinical practice. There are many studies currently underway that are in the first stage, trying to build these optimal rules (in particular, data from Sequential Multiple Assignment Randomized Trials serving as “first-stage data”). However, to my knowledge, the efficacy of the second stage is limited to hypothetical questions and scenarios (counterfactual outcomes) used in causal inference and statistics (e.g., what is the average outcome if everyone had followed their optimal rule?). There is no current “direct” support for implementing optimal treatment rules in practice; however, currently there is a lot of thought going into how to use machine learning algorithms in clinical practice, generally (see: https://jamanetwork.com/journals/jama/fullarticle/2675024). Utilizing individualized treatment rules could yield improved health outcomes, because people would be given the treatment they benefit from the most, as opposed to the treatment that was shown to benefit overall in a population (however, I also understand the discomfort and resistance that may accompany implementing these algorithms in practice). Further, estimating optimal treatment rules with resource constraints allows us to assign an individualized treatment, realizing that we only have a finite amount of resources, which could lower overall costs. For example, if a treatment is very expensive, we can tell the optimal treatment algorithm that only X% of people are allowed to receive the intervention. This is helpful, for example, in the extreme case where the optimal rule says to give everyone treatment.
Thanks, Lina, for the summary of this research topic. This is an issue I have been thinking about as well but haven't done any deep level of reading on it. There is clear gap in demonstrating the efficacy of such decision rules/treatment strategies/etc. that are hypothesized to be optimized for individuals with a certain set of characteristics, but I'm really interested in the "adoption" of such approaches. Personalized approaches are a high priority for the NIH, but given that we have been slow to implement "traditional" evidence-based approaches where needed, I'm wondering how best to avoid the challenges of the past. Further, another issue is the applicability of certain individualized treatments as they would, to some degree, always have an element of clinical judgement to them. A great benefit could be the reduction of any biases when prescribing or adapting a treatment. Looking forward to future discussions on this.
John S
Indeed as JS says, very interesting work. I think you may need to do some thinking about the gap itself. The approaches you are focused on, it seems, might be a strategy for overcoming a gap or barrier, but in your case, you may need to work backwards to identify the gap of interest. One example could be taken from ADAPT - the intervention = ART; the gap = non adherence or non retention; the intervention = a strategy which depends on how you respond to initial exposure to the strategy. There could be others, but first focus on finding that intervention that isn't used...
Hi Lina,
I think your research is very interesting. I might be wrong but I thought of it as a kind of an "artificial" precision medicine. My work is in the area of brain tumors and I definitely can appreciate the value of this implementation since, choosing a treatment is something that the clinicians struggle with on a daily basis. As a junior researcher in this field I also realize the limitations that "conventional methods" can have in our populations, since we work with very small Ns, the treatment options available are limited and it's difficult to generalize results. But I think that methods like the one you are proposing would be very helpful in these kind of settings.
What are the populations that these algorithms have tested?
Have these methods incorporate precision medicine outcomes (i.e. gene mapping results or genetic profiles)?
I look forward to hear more from this research.