MEET DR. AJAY PERUMBETI | OCTOBER 27TH, 2023 

1. First, for those who may not know you. Please share a little background about yourself and your journey as a pediatric hematologist/oncologist.

• Dr. Perumbeti’s journey started off during his pediatric residency, where he spent time abroad in an NGO for a year and started a fellowship in hematology and oncology at the Children’s Hospital in Los Angeles. There he worked at a research lab, where he gained a mentor, and later joined the faculty department as an instructor. Passionate about learning about sickle cell disease and understanding the clinical aspect of the disease, Dr. Perumbeti started a transfusion medicine fellowship. After finishing the fellowship, Dr. Perumbeti became the director of transfusion medicine at Children’s Hospital Los Angeles and became interested in computational medicine, specifically dashboarding, clinical decision support, etc. After working on basic data science and programming fundamentals important in computational medicine, Dr. Perumbeti joined the informatics fellowship at the University of Arizona College of Medicine-Phoenix and Banner Health Systems. 

2. I read in the recent interview with the UA College of Medicine-Phoenix that you are using machine learning models to study iron deficiency. What have been some of your findings?

Over a billion people have iron deficiency globally, and it is the “most common nutritional deficiency,” says Dr. Ajay Perumbeti. Machine learning models can help clinicians understand if a patient is iron-depleted (deficient in iron) prior to developing anemia, preventing morbidity and making it easier to treat. Machine learning algorithms can pick up subtle patterns in commonly found blood counts and electronic health record features, that can predict iron deficiency that are difficult to discern by humans.

3. What has your experience been like using machine learning? 

“Machine learning is an area of research in medicine that is just starting to hit the clinics,” says Dr. Perumbeti. Companies such as IBM Watson, Electronic Health Record (EHR) vendors, Amazon Web Services (AWS), etc are using machine learning algorithms for clinical decision support in medicine. However, according to Dr. Perumbeti, “the biggest thing for a physician would be to evaluate those products and see if they are accurate and contain bias”. Dr. Perumbeti even states, that “these products that implement technology have a shelf life and will need to be adapted over time”. He also states, that “machine learning algorithms have a risk of failure, so they are not always correct”. 

4. What do you think can be done to bridge better the understanding between current medical practice and emerging technologies?

The obvious answer would be more training and education. However, a great tool to understand current medical practice and emerging technologies would be Model Cards. Model Cards are one age summaries of artificial intelligence (AI) algorithms; it tells you what the algorithm is, explain the context in which it was trained, informs you of bias evaluations, and tells you when you should and shouldn’t use the algorithm.  Another hot artificial intelligence category (generative AI) is rapidly being adopted because it can automate content creation.  The most popular, but one of many products is chatGPT, that implements large-language models (LLMs) that use natural language processing methodology called transformers and attention mechanisms. ChatGPT performs word prediction, and writes in a compelling way, but can confabulate and convey misinformation.  The fact that LLMs are so persuasive indicates perhaps that  “human language is generally very predictable”, but is not an indicator if it is telling the truth.  In fact, the “ChatGPT and Artificial Intelligence in Higher Education: Quick Start Guide” from IESALC suggests that, the flow chart below can help us understand when it would be safe to use chatGPT. 

5. In a recent interview with the UA College of Medicine media team, you shared that “Bad information is just as bad as no information,”  - can you share more about your perspectives on this important point?

Knowing that the human brain has limited bandwidth, we have a certain capacity to learn, take in information, and make decisions; because of increased medical knowledge which is growing exponentially, we need tools to use this knowledge to help people.  At the same time, these tools come with intentional or unintentional bias which can lead to worse decisions.  Once we automate decisions, we may place too much trust in the decision (automation bias).

Dr. Perumbeti states that, “We need these tools, and they can reduce human burdens that can be associated with growing amounts of information, can provider physician burnout, being overwhelmed, being exhausted”, but we need systems to evaluate, deploy, and monitor if they are working as intended. 

6. For those who are hesitant to integrate or trust AI in healthcare, how might you address common concerns? 

Having a process of evaluation will give the benefits and limitations of artificial intelligence (AI). The current types of AI in medicine is fairly new, so it is “susceptible to fail and can contain bias”. Having the tool of technology can hopefully ease the burden of ability to do work and the care that physicians give to their patients. 

7. You are currently a Clinical Informatics Fellow at the University College of Medicine-Phoenix.  Can you share a little about why you chose to pursue this training and what you are hoping to do with clinical informatics approaches in the future?

Dr. Perumbeti describes his interest in clinical informatics as an “obsession of addressing issues in medicine to make clinical decisions and improve outcomes for patients”. Dr. Perumbeti states that “there is a great promise to incorporate real-world data from patients including biometric signals, genomic and other omics data including patient environment (exposome), to improve patient outcomes and quality of life.”  

Dr. Perumbeti also explained how the concept of digital twins fascinates him in the field of clinical informatics. The concept of digital twins is ‌a virtual version of yourself that contains all the information required to understand yourself from a biological perspective, having the data to interrogate a version of someone. For example, we could test a specific drug on a digital twin without ‌testing the drug. 

8. What advice would you give current medical students and residents who might want to pursue a career like yours? 

For those who want to pursue a career that implements technology in medicine, it will be critical to have medical domain expertise, and physicians, nurses, and other healthcare providers will be required to implement technology in a way that helps patients. Another aspect would be to develop experience with developing technology as products, and its foundations in computer science, data science, and mathematics. One aspect of understanding the field of medicine and technology could be understanding what an implementation scientist does. Being an implementation scientist allows one to look at how to integrate the technology and using process and evaluation to ensure it is effective. 

Sources:

IESALC. "ChatGPT and Artificial Intelligence in Higher Education: Quick Start Guide." IESALC, 2023, https://www.iesalc.unesco.org/wp-content/uploads/2023/04/ChatGPT-and-Artificial-Intelligence-in-higher-education-Quick-Start-guide_EN_FINAL.pdf.

By: Dr. Ajay Perumbeti and Himanshi Kapoor