Part eleven of our AI Researcher Profile series invites Professor Tengfei Ma, Assistant Professor at the Department of Biomedical Informatics, with affiliation to the Department of Computer Science and Applied Mathematics & Statistics at Stony Brook University, to discuss his research interests and knowledge surrounding healthcare, AI, and machine learning.

AI Institute: You’ve had an exciting career before you started teaching at Stony Brook — from obtaining your Ph.D. at the University of Tokyo to working at IBM. What was your research focus at IBM, and what inspired you to switch to academia?

Professor Ma: My research at IBM was quite similar to what I'm doing at Stony Brook. To begin with, my Ph.D. study was about Natural Language Processing and Bayesian Learning. Later, when I started working at IBM, I began exploring Deep Graph Learning because, at the time, Graph Learning was becoming more beneficial for a lot of industries. I started to look into how Machine Learning and Deep Learning can be used in graphs. At the same time, I was collaborating with the Healthcare Research team at IBM and working on applications of AI in Healthcare.
I moved to Stony Brook after that, because it allowed me to work on my research more consistently, and also because the institute has its own hospitals, so I had access to more real-world data and could directly collaborate with clinicians, especially when I needed to validate the proficiency of my AI models.

AI: You are an Assistant Professor at the Department of Biomedical Informatics, with affiliation to the Department of Computer Science and Applied Mathematics & Statistics. Tell me about your research interests in healthcare, machine learning, and natural language processing (NLP).

TM: My work explores the applications of AI in Healthcare. So one of the areas I’m focusing on is EHR (Electronic Health Record) analysis, which covers a lot of different aspects of healthcare. For example, consider patient history, which you can chart and analyze to perform diagnosis, or make predictions. Using it, you can tell if, in the next year, a patient will have some kind of disease, and determine the risk of a patient's condition. Another example is how, when given a medical image, you can detect whether a patient has cancer. Or you can summarize the patient's history and automatically generate a discharge summary. All these applications have the potential to reduce a clinician’s workload.
Another usage of this technology is in insurance. Clinicians need to generate medical codes for billing needs, so the insurance company can easily gather all the required patient information in time. We can easily create such applications using EHR analysis and Machine Learning.
Now, my major research is about Graph Learning — a branch of Machine Learning that focuses on the analysis of data given to us in the form of a graph, which happens more often than you’d think, because graphs represent real-world data (like social networks) more effectively. Graphs have a lot of special properties, and I’m looking for ways to utilize them for Healthcare. For example, if an AI doctor has a new patient, generally it will only use that patient's medical history to do their analysis. But that’s not how doctors’ minds work. Their experiences with other cases allow them to draw predictions for the new patient. So, when the AI doctor knows this patient is similar to another patient, it can build a similarity graph to connect their histories and use graph learning to make more accurate predictions.

AI: What are some of the most exciting developments we’re seeing in AI and biomedical informatics today?

TM: The most exciting developments of AI these days are Large Language Models and Generative AI like ChatGPT. Their impact extends to so many fields, including Healthcare and Biomedical Informatics.
For example, I recently learned about a project on pathology image analysis, where they gathered a lot of images from pathology labs and built a foundation model, whose purpose is to perform a variety of general tasks, including conversing in natural language and generating text and images. Now, researchers don’t need to build a pathology image analysis tool from scratch to use them for a specific application. They can simply start with this foundational model, and build things up from there!
Similar things are also happening for graph learning, where we’re seeing more and more foundation models being built, which can now be connected to other forms of information, like text, to generate detailed descriptions of graphs.
Now, because anyone with relevant knowledge and tools can do the analysis, you can see the impact of these kinds of foundational models everywhere.

AI: You talked about Graph Learning before. Could you tell us more about Graph Neural Networks and what you’re doing with them?

TM: I work on a lot of different aspects of Graph Neural Networks. When I started out, my focus was more on the Scalable Learning of Graph Neural Networks. So, for example, if you have a large graph, how do you lower the amount of computer memory needed to store it? That’s a big issue. Also, the time required to analyze and infer from the graph is quite huge. And then there’s the need to reduce the training time for the AI model.
Later on, I started to look into how we can expand Graph Learning into Dynamic Graphs, and how to make neural networks more expressive, how to combine Graph Neural Networks with Geometry and Topology.
I also work on theoretical aspects of graph learning models. For example, recently, we submitted one paper about graph transformers, proving why they’re better than traditional graph neural networks. Now, my lab is working on using graphs to make AI easier to interpret and combining graphs with Large Language Models.
All of this is very interesting, because the applications of graph learning extend to so many domains, and especially Healthcare, for applications in drug discovery and drug property prediction, or for enhancing diagnoses and medication recommendations for patients.

AI: You worked on predicting the closure rate of wounds. What is that about?

TM: So, while working at IBM, I was the Co-Principal Investigator on a project for DARPA, a government agency. We worked in collaboration with Columbia and some other universities, and our main goal was to develop a solution for accelerating wound healing.
We started by building devices that could collect sensory information, and used that data to perform a time-series analysis. For example, we could look at the sensory data from a wound, and analyze it to predict its current stage and the speed at which it was healing. Then, based on the results, we could deliver the required drugs in time to speed up their recovery.

AI: You’re also interested in NLP research and generating good document representations and summaries, like Slide4N. This can be used to create presentation slides from computational notebooks with human-AI collaboration. Are you working on other such projects at the moment?

TM: That research is not my main job. It is a product from collaboration with the University of Waterloo, and I still have other ongoing projects with them. There are two aspects to Slide4N — it’s an intersection between HCI (Human Computer Interaction) and NLP.
On the HCI side of things, we recently submitted another paper on using Large Language Models to design posts for social media influencers. You could give this model a text description of what you need, and it will come up with relevant images and captions for you. So, yeah, we designed a very complex end interactive tool for that, and it was very interesting.
Regarding the NLP part, I also worked on other related projects for code summarization and code clone detection. So, for example, I recently worked on a project related to machine-generated code detection. If you look at a snippet of code and want to find out if it was written by a human or a machine, how do you do that? While those projects are not my primary focus, they are very exciting to me.

AI: What are some of the greatest challenges we’re facing in AI and biomedical research today?

TM: Although Artificial Intelligence is thriving and we have a lot of LLMs and foundational models impacting a number of lives, I still think we need to improve AI’s capabilities. For example, you’ll often find that these Large Language Models start to hallucinate and start saying things that aren’t factual. This is an issue we need to resolve.
Another big issue is the lack of data, which is stopping us from scaling AI. In the near future, we will have used all the information we can to train these systems, creating a bottleneck that won’t allow us to further improve AI. How do we solve this problem? That’s a big question, especially for the healthcare industry, because we don’t have access to public datasets. What we do have is private data, but we cannot share this with fellow researchers. And this challenge goes hand in hand with another — privacy and security, which we also need to care about.

AI: This is a consistent challenge we face in this industry, providing data privacy and security. Researchers admit that this is a problem and that we need to do something about it. But are we taking action?

TM: I'm not an expert on this topic, but I know that in Healthcare, people have started to use a concept called ‘Federated Learning’ to train AI. For example, if multiple hospitals don't want to share their data with each other, but want to use the data together to train a common model, they can use this technique to get past this issue. And that's just one of the things that I know of, but of course, there are many other initiatives.

AI: What is the future of artificial intelligence and biomedical informatics? Do you have specific goals that you are working towards?

TM: We need AI to be easier to interpret, especially when it comes to healthcare and biomedical research. One way to do this is when the system performs some kind of diagnosis, it doesn’t just share its predictions but also explains how it arrived at that conclusion. As of now, this feels like a long-term goal, but with AI evolving so quickly, it might happen earlier than we expect.
As for me, I have also started to look into enhancing AI’s reasoning capabilities using graphs or neuro-symbolic models. The goal is to make AI more powerful, and hopefully, more impactful and beneficial to patients’ lives.

Communications Assistant
Ankita Nagpal