Associate Professor of Orthopaedic Surgery and Clinical Anatomy, Mayo Clinic
MD, MS, Clinical Translational Science, Mayo Clinic Graduate School
MS, Orthopaedic Science, Mayo Clinic Graduate School
BS, Neurobiology, University of Washington
@Mayo_OSAIL
Who have been your mentors?
Robert Trousdale, MD (Ortho), Rafael Sierra, MD (Ortho), Michael Taunton, MD (Ortho), and Bradley Erickson, MD, PhD (AI, Radiology).
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 took a keen interest in AI around 2019, seeing that it was a technology coming of age that could help us solve problems in new ways, particularly through leveraging large and well curated datasets that we are fortunate to have at Mayo Clinic. My first project was a proof-of-concept idea to see if we could create an automated calculator to determine hip arthroplasty acetabular component position (inclination and anteversion angles) on x-rays. This has important implications for dislocation risk and postoperative surveillance. The calculator proved quite accurate to within 1 degree of expert human annotators and performs the analysis in about 5 seconds. We leveraged that speed to annotate all the patients radiographs in our Total Joint Registry dating back to 2000 (roughly 30,000 THA patients) and did this in 1 day.
What is/are your current interest(s) in the field of AI/ machine learning (ML)/ deep learning (DL) for orthopedic applications?
Our Orthopedic Surgery Artificial Intelligence Laboratory (OSAIL) at Mayo Clinic is working on several areas of interest across all subdisciplines. We work heavily with imaging data as that is the cornerstone of orthopedic evaluation, but we also leverage electronic medical records data to get demographic and comorbidity information as well as to analyze the free text of clinical notes with large language models (LLMs; i.e. think ChatGPT) to garner data that is otherwise uncaptured. Using these data sources, we are creating imaging classifiers and automated annotation tools, patients-specific preoperative risk calculators, and automated clinical registry curation with LLMs. I am perhaps most excited about using generative AI in the imaging space where we are creating tools to predict what a patient’s future x-rays will look like over time as well as taking a single x-ray and being able to rotate it in space with a synthetic project, all the way up to creating 3D models from single x-rays.
What has been your most unexpected result when using AI/ML/DL for orthopedic applications?
I have been most surprised by DL models having the capability to evaluate a single preoperative x-ray and make better patient-specific risk predictions for complications such as hip replacement dislocation than a multimodal model based on known risk factors. These algorithms are very powerful and enabling new insight into what drives pathology.
What advantages and challenges are associated with using AI/ML/DL (in general or for orthopedics) at the present time?
There are many challenges, but I think the biggest is finding datasets that are large enough, well cleaned and curated to train models. These models are very data hungry, and medical data is relatively sparse compared to natural images for example. Most centers do not have millions of x-ray or chart examples to train models on, unlike how tech companies can instantly evaluate billions of pictures on the internet to train a model.
Is there a specific barrier that once addressed might help expand the use of AI/ML/DL in fracture repair research? How might your current AI methodologies be applied to improve fracture research and clinical care?
Similar to my comment above, we need to get large, clean, and well curated datasets. One way to do this is share data across centers, but this is logistically very challenging. However, if this can be accomplished, the resultant models will be much more powerful and generalizable. I believe AI will eventually be able to analyze a patient’s fracture and determine a personalized phenotype to inform individualized treatment and prognosticate fracture repair outcomes.
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 first step is to get a good understanding of what data you have access to and how much manual annotation needs to be done to make the data “AI ready” to train a model. This is the most laborious and underestimated step by most investigators. Make sure you work with a knowledgeable data scientist as that is critical. For beginners, I recommend courses on Coursera to get started and then advance from there with online resources that are becoming more abundant.
Leave A Comment