I frequently receive emails from potential M.D.-Ph.D. students seeking advice on career decisions. Below is a summary of the common questions about pursuing research in machine learning for healthcare and insights into my career path. I hope you find this helpful, and please feel free to email me hyewonjeong.contact@gmail.com if you have further questions (I might not be able to respond quickly but will try hard to respond quickly).
This page is still under construction. I have received similar questions from students over the years, so I am gradually collecting my answers here. These are based on my personal experience rather than universal advice.
My path has been somewhat interdisciplinary. I studied biotechnology (synthetic biology) and neuroscience as an undergraduate, went to medical school, and later pursued graduate training in computer science before eventually focusing on machine learning for healthcare.
I did not start with a perfectly defined plan to become an “ML for healthcare researcher.” My interests became narrower as I gained more exposure to different types of problems. Medicine made me interested in clinical questions and how decisions are made in healthcare, while computer science gave me tools to study those questions at scale.
I chose to pursue a Ph.D. because I realized that I enjoyed research itself: identifying problems, developing methods to study them, and trying to understand why something works or fails. I wanted the freedom and training to pursue research questions more independently rather than only applying existing tools.
I narrowed my interests mostly by actually working on different problems rather than deciding everything in advance. Early in training, I think there is significant value in exploring. Working across different areas helped me understand what kinds of questions I enjoy, what I am good at, and what I do not want to spend my career doing.
At some point, however, breadth alone stops being enough. To do meaningful research, you eventually need enough depth to understand what has already been tried, what the important unresolved questions are, and what constitutes a genuinely new contribution.
I do not think there is a universal point at which everyone should specialize. My preference would be to explore relatively broadly early on, but once you find a set of problems that you repeatedly return to, begin developing deeper expertise around them. Your research topic can continue to evolve even after that.
My medical background has been particularly useful for problem formulation and interpretation. In healthcare ML, it is relatively easy to obtain a statistically interesting result that does not necessarily make clinical sense. Clinical knowledge helps me ask whether a cohort is appropriately defined, whether a label actually represents the clinical concept we claim it represents, whether an observed association is plausible, and whether a model may simply be exploiting artifacts of the healthcare process.
It is also useful when a result looks unexpectedly good or unexpectedly bad. Instead of treating performance metrics as the endpoint, I tend to ask what could be happening clinically or in the data-generating process. That said, I do not think you need an MD to learn how to ask these questions.
While having an MD can certainly help in understanding the domain, students without an MD can also conduct research effectively. They can collaborate with healthcare professionals and gain a solid understanding of the research problems, making an MD not necessarily essential.
I have received this question many times, and I am probably not the best person to answer it in terms of whether it is objectively “worth it.”
I did not choose my training path by calculating whether an MD/PhD would maximize my career return. I wanted to study medicine, I wanted to do research, and this was one of the paths that allowed me to pursue both.
An MD/PhD is an unusually long training pathway. If someone primarily wants to develop ML methods for healthcare, I would not recommend pursuing an MD solely because they think the credential is required for healthcare AI research. It is not.
On the other hand, if someone genuinely wants clinical training and wants medicine itself to remain an important part of their career, then the calculation is different. Ultimately, I think the decision should depend much more on what kind of work and career you actually want than on whether the degrees themselves are “worth it.”
People also say given the time it might not be worth it for some people and it will never be worth it.
For someone coming from computer science, I would not recommend trying to learn “all of medicine.” That is neither realistic nor particularly useful. Instead, I would learn enough about the specific disease, clinical workflow, measurement process, and decision being studied to understand:
what the real clinical problem is;
how and why the data were generated;
what important confounders or sources of bias might exist;
what constitutes a meaningful endpoint; and
whether the model's behavior makes clinical sense.
The level of medical knowledge required will depend heavily on the research question.
I have done both, but increasingly I prefer research in which there is a real healthcare problem first. I generally find it less interesting to start with a new ML method and then search for a medical dataset on which to apply it. A technically sophisticated model is not necessarily useful simply because it can be applied to healthcare data.
At the same time, I do care about methodological questions. Some of my work is motivated by broader ML problems such as robustness, fairness, multimodal learning, or learning from imperfect clinical data.
The research I find most compelling is usually at the intersection: there is a meaningful clinical or scientific problem, and solving it exposes an ML problem that is itself interesting.
In that situation, the healthcare problem motivates the methodology rather than serving merely as an application domain.
Answer: As I am not a professor directly involved in selecting PhD students, it's difficult for me to provide a definitive answer. However, having research experience is likely to be important.