Orthopedic field awaits impact of artificial intelligence
Upkeep of artificial intelligence necessary for accurate, effective patient care
Click Here to Manage Email Alerts
Since the 1950s when the term artificial intelligence was coined, its application and use has increased through rapid technological advances and has found their way into the health care sector, including orthopedics.
Source: Scarlett Ames
A study published in 2018 showed the amount of orthopedic literature on machine learning, which is one type of artificial intelligence (AI), had an approximate tenfold increase since 2010, with the most frequently applied machine learning algorithms found in spine pathology, osteoarthritis detection and prediction, and imaging of bone and cartilage.
“I think there has definitely been an increase in our understanding but also our attraction or fascination with how [artificial intelligence] may shift care in orthopedics going forward,” Atul F. Kamath, MD, director of the Hip Preservation Center, staff in the Orthopedic and Rheumatologic Institute and professor of orthopedic surgery at Cleveland Clinic, told Orthopedics Today. “I think qualitatively, whether you are a lay person or someone in the medical field, you know artificial intelligence is integrated into multiple facets of daily life with autonomous cars and Siri, but also has merged into the medical world with projects like IBM Watson and Google platforms.”
An increase in larger datasets along with the convergence of cloud-based computing and graphical processing units (GPUs) with other areas of technology have allowed AI to become what it is today, according to Joseph H. Schwab, MD, chief of spine surgery and associate professor of orthopedic surgery at Harvard Medical School and Massachusetts General Hospital. He said convolutional neural networks and deep learning algorithms have been used in image analysis while predictive models using machine learning algorithms have been used to identify outcomes of surgery or treatment.
“Whereas before you might use linear regression, things have evolved into more people using machine learning algorithms and many of those machine learning algorithms evaluate the data even if it were non-linear,” Schwab said. “Linear regression, of course, the data is presumed to be linear, but in real life most data are not truly linear and so using algorithms that allow that flexibility can help with prediction models.”
AI across pathways of care
By using registry data, including demographic and comorbidity information, preoperative disability mental health scores and radiographic measurements, Christopher P. Ames, MD, of University of California, San Francisco (UCSF), noted AI prediction models could be used to assist in real-time decision-making for patient care and treatment by preoperatively predicting major postoperative complications, as well as risk of reoperation and of readmission based on patient and surgical factors. Models can also be created to predict the specific type of benefit of a surgical procedure to an individual patient including precise patient-weighted priorities, such as ability to return to work or decreased pain medication usage, according to Ames.
“Immediately at the point of care, we can give [patients] the risk-benefit profile that is accurate and has been validated, and we can adjust the surgical variables, as well as show patients in real-time the risk/outcome benefit of preoperative physical conditioning or smoking cessation” Ames told Orthopedics Today.
He said, “If a smaller operation has a similar benefit profile to a larger operation but lower complication profile, that assists us in real time in selecting that procedure for the patient.”
Ames noted these predictive models also can be used to predict cost per procedure, which may help institutions protect their bottom line and enter into risk-sharing relationships with payers.
“We can predict with about 75% accuracy the 90-day cost, which is important for bundled payments, and we can predict with 90% accuracy whether patients are likely to reach the Medicare allowable cost,” Ames said. “This becomes important for hospitals and surgeons to be able to protect the price point,” he said.
Intraoperatively, AI may lead to more accurate surgical decision-making at the time of surgery and improved patient outcomes by allowing surgeons to tailor each surgical experience to a particular patient, Kamath said.
“A knee replacement is still a knee replacement in the sense that the same conventional implants are being put in, but how do we take each surgery and make it more like a fingerprint? How do we tailor each surgical intervention to the individual patient? I may want to balance the soft tissues one way for this particular patient and for another patient I need to address the bony deformity in a different way,” Kamath said.
Augmented or virtual reality types of AI may also be beneficial with rehabilitation protocols postoperatively by using a machine learning algorithm to improve patient care, according to Schwab, who noted this could be especially beneficial for patients who live in rural areas.
“The machine learning algorithm would be picking up how the patient is doing or if they are not doing well and providing feedback not only to the game so that the game can be adapted, but also to their treating team so that they know the patient is having trouble,” Schwab said.
Use in traumatic injury treatment
However, machine learning is not the only area where AI may provide benefits to patients with musculoskeletal injuries.
Stephen F. Badylak, DVM, PhD, MD, professor of surgery at McGowan Institute for Regenerative Medicine at University of Pittsburgh, said AI may be able to identify whether tissues that are healing from a traumatic injury will have a good or bad outcome based on certain biomarkers.
“For example, the more scar tissue and adhesions you get, the worse the clinical outcome is going to be,” Badylak said. “If we knew early on that certain things were starting to go down the wrong pathway, we could intervene either surgically or in other ways that could change the direction of wound healing,” he said.
Once the healing biomarkers are identified, Badylak noted AI, in the form of smart bandages, can be developed that, when used on wounds following a traditional repair procedure, can indicate in which direction the wound is healing and provide guidance on how surgeons can intervene, and when. Currently, Badylak and his colleagues are using AI to develop methods “that will decrease the time to functional healing of injured musculoskeletal tissue by 50%.”
“In other words, someone comes in with a traumatic injury, they have a fractured femur and they lost 50% of their quadriceps muscle. That is a pretty serious injury. Right now the expectation is we are going to do the best we can to salvage this and we are going to get to a point where we have scar tissue with a lot of morbidity and limited function,” Badylak told Orthopedics Today. “Well, the AI tells us that does not have to be the case. We can not only restore most of the functional tissue that was gone, but we could do it in half of the time if we only knew when to intervene and how to intervene both surgically and medically.”
Feedback aids learning
AI may be used as a learning tool for orthopedic surgeons through its feedback mechanism, which Kamath noted can learn and develop over time.
Surgeons also can review benchmark models to compare their outcomes and complication rates to those of centers of excellence and learn which patients may benefit the most from which surgical procedures, Ames said.
“What if the surgeon sees a case, he runs it through the model and it says projected complication rate 70%?” he said. “That might be a surgery they decide not to do or it might be a surgery that they defer to a center of excellence, to send it to one of the better sites in the U.S. to look at and figure out whether the risk-benefit profile will be better at that site.”
Schwab said a GPU can be used in addition to machine learning to help teach residents and medical students on different surgical techniques.
“You could have a lab or a computer game that, on the one hand, could simply train your decision-making using machine learning,” Schwab told Orthopedics Today. “Some people have more advanced understanding of a problem. The machine learning algorithm could adapt to that and ask them more advanced questions and vice versa, but you can also use that for computer simulation. So, you are doing technical work and having direct feedback from a machine learning algorithm.”
Barriers of AI
Despite the current interest in AI, Mark Alan Fontana, PhD, senior director of data science at Center for the Advancement of Value and Musculoskeletal Care at Hospital for Special Surgery, noted more research is needed before AI can make a bigger impact in orthopedics. This lack of research may be due to a lag in the use of technology in health care compared with other industries, he said.
“I think health care in general tends to lag behind other industries in terms of using tech and adopting new methods,” Fontana told Orthopedics Today. “Part of that is because it is inherently conservative ... by nature, given the stakes are high. We are dealing with life and death. The security issues are also paramount, all the HIPAA regulations. It makes data more challenging for researchers and others developing these things to share across organizations and so data tends to be more siloed.”
One barrier of the use of AI in orthopedics is the learning models and algorithms are only as good or efficient as the data that are put in, according to Kamath.
“We have to constantly evaluate and make sure this is a safe introduction of technology and make sure that the hype associated with this is matched to the reality of what it can give at that point in time,” Kamath said.
Although the data that AI and predictive models compile can be a good source of hypothesis generation for research, Fontana noted these models can detect patterns in prior data and apply these patterns to newly input data, which may lead to predictive models giving advice that parrots societal and social biases within the data.
“It highlights a necessity of testing models not only for their accuracy, not only whether they work in the real world ... but also if they are equally effective on all types of patients or if they are not equally effective, quantifying that and properly caveating that as they are being used,” Fontana said. “The last thing you would want to do is put something in place that makes the situation worse for groups of people in our society that are already discriminated against,” he said.
Furthermore, predictive models “tend to drift” or do worse as reality changes over time. They therefore require continuous upkeep, which may not always be possible depending on the organization. In addition, these models are not one-size fits all and need careful curation from hospital to hospital, according to Fontana.
The long set of operating instructions associated with predictive models “is time consuming from a resource standpoint and from an understanding standpoint,” he said.
Because not every hospital will have the resources to run and upkeep these models, Badylak said the models must be developed in an understandable and user-friendly way, especially in light of the current barriers to implementing AI as a standard of care. He said regardless of whether they are user-friendly, the more advanced AI designed for treatment of musculoskeletal injuries may be disadvantaged by regulatory barriers, as well as funding issues.
“The only way you are going to [obtain] third-party reimbursement is if you can show that, in the long run, this is more cost effective,” Badylak said. “That takes years to do because the cost benefits are going to be realized over years,” he said.
‘Keep an open mind’
Although AI is still in its infancy, experts with whom Orthopedics Today spoke believe AI will be beneficial in orthopedics as it evolves.
It may help with answering difficult questions on the individual patient level and incorporate payment models, costs and insurance, Kamath said.
“These are all big questions that require lots of data and I think as we go forward in orthopedics, especially in hip and knee care, which is a big player in terms of the amount of dollars that are spent on hip and knee replacement, I think optimizing AI in these ways would be helpful, not just from a clinical standpoint, but also from a health systems and resource utilization standpoint,” Kamath said.
AI may also help orthopedic surgeons advance and possibly improve the way they practice medicine, he said.
“It will help us question older dogma or the way we practice medicine for different reasons. I think the medical community can take these emerging platforms and harness that and empower them to deliver care in a better value-based fashion,” Kamath said. “I would say surgeons and clinicians should be at the front edge of innovating these pathways in patient-centric models of care.”
Although Badylak said orthopedic surgeons should keep an open mind regarding AI use in their practices, he also advises them to educate themselves and not believe everything they hear.
“If we are feeding the wrong information into a device that works on AI, then we are not going to get the information out of it that we want,” Badylak said. “In addition to keeping an open mind, we have to put forth the effort to understand how to take full advantage of what is being developed, otherwise we are going to either not make any advances or even have worse outcomes because we have provided the wrong information and expect a machine to fix it for us, and that does not happen.”
- References:
- Ames CP, et al. Spine (Phila Pa 1976). 2019;doi:10.1097/BRS.0000000000002974.
- Ames CP, et al. Spine (Phila Pa 1976). 2020;doi:10.1097/BRS.0000000000003242.
- Cabitza F, et al. Front Bioeng Biotechnol. 2018;doi:10.3389/fbioe.2018.00075.
- Helm JM, et al. Curr Rev Musculoskeletal Med. 2020;doi:10.1007/s12178-020-09600-8.
- Machine learning for the practicing surgeon: SORG. Available at: www.sorg-ai.com. Accessed June 11, 2020.
- Myers TG, et al. J Bone Joint Surg Am. 2020;doi:10.2106/JBJS.19.01128.
- For more information:
- Christopher P. Ames, MD, can be reached at 400 Parnassus Ave., Third Fl., San Francisco, CA 94143; email: christopher.ames@ucsf.edu.
- Stephen F. Badylak, DVM, PhD, MD,can be reached at 450 Technology Dr., Suite 300, Pittsburgh, PA 15219; email: badylaks@upmc.edu.
- Mark Alan Fontana, PhD, can be reached at 535 East 70th St., New York, NY 10021; email: ironsm@hss.edu.
- Atul F. Kamath, MD,can be reached at 9500 Euclid Ave., Cleveland, OH 44195; email: elyk@ccf.org.
- Joseph H. Schwab, MD,can be reached at 55 Fruit St., Boston, MA 02114; email: tjanos@partners.org.
Collapse
Click Here to Manage Email Alerts