Clinical Chemistry - Podcast
Thomas J S Durant, Seung Joo Lee, Sarah N Dudgeon, Elizabeth Knight, Brent Nelson, H Patrick Young, Lucila Ohno-Machado, Wade L Schulz. Quantum machine learning and data re-uploading: Evaluation on benchmark and laboratory medicine data sets. Clin Chem 2026; 72(4): 451–60.
Dr. Thomas Durant is an Associate Professor of Laboratory Medicine and Biomedical Informatics and Data Science at the Yale School of Medicine and Dr. Wade Schulz is an Associate Professor of Laboratory Medicine and computational healthcare researcher at Yale School of Medicine.
Bob Barrett:
This is a podcast from Clinical Chemistry, a production of the Association for Diagnostics & Laboratory Medicine. I’m Bob Barrett. In the age of big data, machine learning has shown the ability to take complex data sets and draw meaningful insights that would otherwise have gone undetected by other analysis techniques. In the clinical laboratory space, machine learning has already been applied to classify images and predict disease risk, but unfortunately, current algorithms have important limitations that prevent more widespread implementation.
As an alternative approach, quantum computers offer the potential to operate more efficiently than classical computers and may extend machine learning even further into a variety of hospital environments. Indeed, quantum machine learning has shown promising results using relatively simple data sets, but it hasn’t been widely evaluated in large, complex data sets frequently encountered in healthcare settings.
A new research article in the April 2026 issue of Clinical Chemistry tests several machine learning algorithms head-to-head using data sets of increasing complexity. Today, we’ll speak with two of the article’s authors. Dr. Thomas Durant is an Associate Professor of Laboratory Medicine and Biomedical Informatics and Data Science at the Yale School of Medicine. His research involves the practical application of novel computing and informatics technologies.
Dr. Wade Schulz is an Associate Professor of Laboratory Medicine and computational healthcare researcher at Yale School of Medicine. Dr. Schulz has research interests in the management of large biomedical data sets and the use of real-world data for predictive modeling.
So, Dr. Schulz, let’s get started with you. What motivated your group to begin investigating the applications of quantum computing and biomedical research and laboratory medicine?
Wade Schulz:
Yeah, thank you for the question. So, this was an area that we’ve started out working on over the last couple of years now. And one of the areas of interest for our group is to assess the application of new technologies to biomedical and clinical informatics.
And quantum computing has really been on this new frontier of how can we work with data and compute to try to accelerate and build better algorithms, build more efficient algorithms, much like we saw with the transition from CPUs to GPUs. And that’s an area that’s definitely taken off over the last really decade plus.
Do these QPUs, or quantum computing units, provide new advantages that we can take advantage of both within biomedical informatics more broadly, as well as laboratory medicine and some of the unique data sets and use cases that we have within the specialty?
And so that’s really triggered off for us a number of different applications, some of this using data from the clinical chemistry lab, others looking at areas of genomics and functional genomics, and one that we’re continuing to explore.
While it’s an early area of computing, for quantum computing in general, one that we’re very interested in, excited about, and one that is really open to a number of different experimental and use cases that need to be assessed to see what its long-term potential might provide to both the biomedical informatics community at large, as well as those of us in laboratory medicine.
Bob Barrett:
So, Dr. Schulz, what potential advantages does quantum computing offer in these fields? How might it extend or enhance capabilities beyond those of classical computing?
Wade Schulz:
Yeah, and that’s an interesting question that we really don’t fully know the answer to yet, and one that makes this field of research very exciting.
So, the pitched potential advantages are one that we can potentially compute on data sets more quickly, that the quantum computing, that quantum processors can do that more efficiently than a CPU or GPU, at least for certain types of math and certain types of tasks. It also has the potential advantage of improving things like the performance of predictive algorithms.
However, within the field, a lot of this is still somewhat theoretical, or where it’s been applied, it’s been applied to relatively small data sets. Part of that is because we’re still figuring out which algorithms work the best on quantum hardware, and others are that the quantum hardware itself is still relatively new to being available broadly outside of the labs that are building the actual computers themselves.
And so, as we’re moving through this, it’s really at that experimental stage of does the promise and theoretical advantages actually play out practically, and can we identify the data sets and applications that have the most potential?
One example of this is that for doing something like searching large databases or trying to do graph-oriented algorithms like community detection may be more efficiently or accurately solved by quantum applications or the quantum processing units and the associated algorithms used on top of them -- but is still very new to see does that actually play out, and how much of an advantage do we get, and in which areas or use cases do we see that offering an advantage, both in those performance aspects as well as in cost, compared to the classical counterparts.
Bob Barrett:
Well, thank you. Dr. Durant, let’s turn to you. What are the key findings from your publication and what do they imply for the field?
Thomas Durant:
Yeah, thanks Bob. I think going through the paper, it’s worth highlighting three main takeaways.
And so, we looked at it in stages, so we’ll go through it in a linear sequence. And the first one being general performance. I think our first question when looking at this was -- and we’re not, full disclosure up front, we’re not expert mathematicians or machine learning experts, we can’t do calculus or complex linear algebra on paper. We can just plug these things in to run them.
But for in our minds, we think of there being these two types of problem sets. Ones that have a linear decision boundary and a non-linear decision boundary. The difference being between the two that with a linear decision boundary, you can picture blue dots and red dots on a two-dimensional graph and you can draw a straight line between them that would separate the two classes. And then a non-linear decision boundary would be one that’s curved, right?
And so, we started off by asking the simple question of, can quantum-based machine learning model solve a problem that has a non-linear decision boundary? And so, for that, we started with a circle data set and compared it against machine learning algorithms like linear regression, linear SVC, that we know were just capable of learning linear decision boundaries. Because that means a lot about a model in general and an algorithm, to say that it can learn a complex decision boundary.
And so, that was one of the first takeaways for us to observe, which was the quantum re-up, the QNN, and the VQC were all seemingly capable of learning a non-linear decision boundary, which is good. That means that they had the potential, if you scale that up into a larger input feature space, meaning more variables in the data set, they might be able to learn on bigger problems. This is all in Table 2 in the paper, by the way.
But if then you look at the quantum re-up algorithm compared to the QNN and VQC, the re-up performed a little bit better on the smaller data sets, and then relative to the QNN and VQC was actually able to get through all of the training on the larger data sets, that being the plasma amino acid and the breast cancer data set, as opposed to others which weren’t able to finish training due to failed convergence or exceeding computational resources. So, that was the second main finding.
The third one was that when we turned all the knobs and switches and levers in the configuration parameters for the QNN re-up algorithm, we saw drastic changes in the performance. And so, performance variability is the third main takeaway. In that, with classical machine learning, you can see -- depending on how you formulate the problem, you can see some swings in performance, but this was much more stark than we are used to in the classical machine learning world.
It can basically swing from -- it looks like the algorithm isn’t learning at all, it’s completely broken, to it’s actually learning pretty decently depending on which switches you flip. Which for us was important because as we’re going through this in the early stages, we had thought we weren’t doing it right, but then we just changed how we were configuring our problem and it started working.
And mind you again, this is all quantum machine learning simulated on classical processing units. So, we may find that if we move to true QPUs in the future, that there’s additional configuration parameters that need to be set. But really, at a high level, the main takeaways were the general performance being able to learn non-linear decision boundaries, re-up doing a little bit better than QNN and VQC, and then the massive performance variability.
Wade Schulz:
Yeah, and really our goal for this was to assess, one, ‘how do these algorithms work, how variable are they?’ as Dr. Durant mentioned, and also looking at that difference between not just a quantum versus classical algorithm, but also different types of quantum algorithms.
We were able to see with some of those configuration tweaks, similar performance but not yet exceeding the classical algorithms. But we’re somewhat limited that the current quantum hardware that we had available at the time, and needing to do some of that in these quantum simulators or quantum emulators, prevents us from including much larger data sets, at least in the quantum computers that have this more circuit-based or general mathematic, general equation type of design.
We are working on some other projects that can take advantage of quantum computers that have far more qubits, or quantum bits, and that has some additional possibilities of showing again improved performance or being able to work with larger data sets.
But again like Dr. Durant mentioned, we were limited in being able to test some of this on actual quantum hardware because the actual physical devices available weren’t large enough, did not have error correction built in to be able to handle the larger volume of data that we would see in a lot of the real-world applications.
Bob Barrett:
Well finally Dr. Durant, looking ahead, what future projects are you planning to pursue using quantum computing? Do you anticipate focusing on gate-based systems, quantum annealing platforms, or both?
Thomas Durant:
Yeah. So, that’s a great question. I think at the moment we’re looking to start doing research and continue to do research in both areas, using both gate-based quantum computing systems and annealing platforms.
And one thing to highlight is that the paper that we just published with Clinical Chemistry, that’s using a gate-based system and those are somewhat analogous to classical computers, which also have gates that operate on binary bits. Whereas the quantum annealing platforms, I think of them almost like an analog computer where there’s no gates, it’s maybe not considered a universal computer in the classical Turing sense.
You have to formulate your problem in a very specific way, in a specific formula, in order to fit it into the quantum processing unit. And so that affords some limitations and restrictions, but it also offers some potentially promising benefits for the optimization that you do during machine learning problems, and finding better solutions potentially.
And so, we’ve seen some success with that early on with pilot studies, looking at community detection-based problems and feature selection problems, and combining the latter with classical machine learning. So, there’s a couple of avenues we’re looking to pursue there.
And the other thing too is that the quantum annealing platforms are a little bit more accessible in terms of getting actual QPU time and doing experiments on actual quantum hardware. So, we’re excited to see what the opportunities are in that area too.
Bob Barrett:
That was Dr. Thomas Durant and Dr. Wade Schulz from Yale School of Medicine in New Haven, Connecticut. They wrote a research article in the April 2026 issue of Clinical Chemistry evaluating the use of quantum machine learning in laboratory medicine. They’ve been our guests in this podcast on that topic. I’m Bob Barrett. Thanks for listening.