Medical Student, David Geffen School of Medicine, UCLA
MPH, Epidemiology, Columbia University
BA, Biology and Computer Science, University of Pennsylvania

Who have been your mentors?
Chamith Rajapakse, Ronald Summers

How did you get involved in research studying the use of Artificial Intelligence (AI) for medical applications? What was your first project using AI in the field of orthopedic research?
I got involved with AI after seeing what can be done with it in my computer science classes. While our classes were focused on toy examples (e.g. classifying images of animals), I saw the potential for deep learning to be applied to the medical world. Around that time, I had begun volunteering in the hospital and saw patients recovering from spinal surgery (often to correct vertebral compression fractures). I thought if there was a way to detect these fractures ahead of time, that could save patients a lot of time and stress, and maybe prevent the need for an invasive surgery in the future. I got connected to Dr. Rajapakse’s lab, where I developed an algorithm to detect vertebral deformities by evaluating morphometric measurements in vertebral bodies from X-rays, CT, and MRI that went on to get published in Bone and Radiology AI and is now patent pending. I firmly believe this tool can help detect and quantify vertebral compression fractures ahead of time. Since then, I’ve worked on that project and multiple others focused on the realm of radiology and opportunity screening.

What is/are your current interest(s) in the field of AI/ML/DL for orthopedic applications?
Currently, I’m working on finding ways to create better hip fracture risk predictors. We rely on DXA scans as one of the components to a FRAX score for determining whether someone needs treatment to avoid hip fractures in the future (due to osteoporosis). However, DXA scanners are not universally available. We are trying to find a way to use pelvic X-Rays instead.

What has been your most unexpected result when using AI/ML/DL for orthopedic applications?
The most unexpected result I’ve seen was while working with another member of Dr. Rajapakses lab, George Asrian. We set out to find predictors of mortality after hip fractures. Using a large admissions database, we found that there were some expected predictors (age, platelet count) but also some unexpected ones such as red blood cell distribution width, mean corpuscular hemoglobin concentration, and calcium levels. That finding just drove home the point that a lot of outcomes are driven not only by the original insult/injury to the bone but also due to many other physiological processes happening at the same time. This work was recently published here.

What advantages and challenges are associated with using AI/ML/DL (in general or for orthopedics) at the present time?
The number one challenge in my mind is dataset curation and data sharing. It remains difficult to assess how models trained using AI perform outside a single institution. There are now some multi-institutional open databases (OAI dataset, MIMIC) but a lot more could be done in terms of data sharing and (once datasets are shared) data standardization.

Is there a specific barrier that once addressed might help expand the use of AI/ML/DL in fracture repair research?
One thing that clinicians and others implementing AI want to see is whether that model will work for patients at their institution. If these models are validated at multiple institutions and if models (or their datasets) are made public for evaluation, that can reduce a lot of the apprehension around trusting whether the results will work for people’s individual patient populations.

What advice would you give investigators who want to include AI/ML/DL into their research program? What learning resources or techniques would you recommend?
The things I think would be helpful for a lab to start doing AI/ML/DL research is to first formulate your question and really figure out if it is necessary to use AI to address it. Secondly, try to gather and standardize all the data you need to address it. It’ll make it much easier to get someone with experience in this area on board if there’s minimal friction to start working with your data and your clinical questions. In terms of resources, I’d recommend Andrew Ng’s AI course via Coursera and simply using tutorials you can find online through blogs and YouTube.

How might your current AI methodologies be applied to improve fracture research and clinical care?
We are already seeing some interest in using the tools we have developed in automating and standardizing measurements of compression fractures and scoliosis in clinical workflows. Down the line, I hope to see a vertebral compression fracture report automatically generated for a radiologist or orthopedic surgeon to review.