Exploring Mayo Clinic’s AI Efforts to Enhance Organ Transplants

August 24, 2023 - Organ transplants represent one of many revolutionary advancements within the field of medicine over the past century. Since the first successful human kidney transplant in 1954, the number of treatment options for once life-threatening or incurable conditions, like kidney disease, has broadened significantly.

Organ transplant-related innovations like tissue typing, genomic sequencing, and antirejection drugs have helped spur progress in this area. But, in the future, clinicians may leverage more advanced approaches to improve patient outcomes.

One possibility currently being researched is bioengineered organs, but some health systems are turning to another alternative that’s received increased attention in the healthcare industry: artificial intelligence (AI).

Mayo Clinic is one of the innovators in the organ transplant AI niche, with efforts underway to use the technology to prevent the need for a transplant, improve donor matching, increase the number of usable organs, prevent organ rejection, and bolster post-transplant care.

Two experts from the health system chatted with HealthITAnalytics via Zoom about the challenges clinicians face in improving transplant outcomes, Mayo’s research into how AI can help, and bumps in the road to deploying these technologies in the clinical setting.

THE CHALLENGES IN IMPROVING TRANSPLANT OUTCOMES

The dynamic nature of transplant medicine is one of the key difficulties clinicians face in improving patient outcomes.

“When we associate outcomes with patients, normally you just have one patient and one outcome,” explained Rohan Goswami, MD, a transplant cardiologist at Mayo Clinic in Florida. “But in patients that receive organ transplants, you actually have two patients now. You have the organ that you recovered, and you have the patient that you're managing, and sometimes they don't like to play well with each other.”

The complexity is even greater in multi-organ transplants, like heart-kidney or heart-liver, which can involve three patients.

Carefully managing and optimizing care for these patients is key, but doing so comes with numerous challenges.

“Our goal is to identify patients and figure out who is high risk for both rejection after heart transplant and who is at higher risk of infection or cancer development,” Goswami said. “Because those are two different timelines, there's a short-term outcome and a long-term outcome.”

The short-term outcome is largely driven by infection, which health systems have studied within and outside the organ transplant field for years. Goswami noted that efforts in the 1990s and 2000s helped significantly to address infection risk by standardizing and creating protocols for activities that can lead to infections, such as catheterization.

Applying a similar approach to organ transplants is much less straightforward.

“There's nothing that's standardized about transplant,” Goswami said. “You can put five hearts in five different people, and you'll have five different outcomes, versus if you put a catheter in somebody and you protocolize the management of that… that's going to be the same for every patient that gets that catheter.”

While establishing protocols and standardization in transplant medicine is challenging, it’s not impossible. Efforts to reduce outcome variability that have already proven effective in this area include regular patient follow-up and effective patient-provider communication, which are critical to involving patients more deeply in their recovery process.

Goswami indicated that doing so helps improve survival while waiting for transplantation and survival up to one year and three years out from transplant. These are some of the metrics used to measure institutions performing transplants.

The changing demographics of the US population also impact transplant outcomes. Patients are living longer than they have in the past, thanks in part to medications, Goswami stated. This has led to them potentially getting sicker and needing transplants later in life, which increases overall patient risk and affects survival for those needing transplants.

Organ transplant risk profiling is designed to help clinicians assess and stratify patients based on a multitude of factors, but there is potential to improve the process, Goswami said.

This is where AI comes in.

MAYO’S WORK IN ORGAN TRANSPLANT AI

When assessing patient risk, transplant providers are tasked with analyzing dozens of variables related to the patient’s medical history, demographics, and likelihood of certain complications, among other factors.

This can make more granular risk stratification difficult, limiting how effectively patients can be categorized as low or high risk.

To address this, Goswami and colleagues have developed an AI tool that uses electrocardiograms to help determine whether patients are at lower or higher risk for organ rejection following a heart transplant.

“[The model is] not something that's available clinic-wide, but it's something that we're using to kind of understand how to define those high-risk patients,” Goswami explained. “Because if you can find the high-risk patients, then you can saturate your clinic visits with those patients, and for the lower-risk patients, especially with telemedicine being so popular, you can do video visits. You don't have to have them come back to the hospital as frequently at the risk of them catching an infection from somebody else that's sick.”

Much of Goswami’s work also centers on using AI and other advanced technologies to diminish the need for a transplant.

One of Goswami’s tools is a heart pump called an Impella, which can be placed in patients to help recover organ function while awaiting a transplant by addressing pulmonary hypertension, a common contributor to organ failure post-transplant.

Research indicates that heart failure patients implanted with the pump achieved significant one-year post-transplant survival rates of up to 95 percent despite risk factors like older age and prolonged wait times for transplants.

The device was intended to help patients survive long enough to receive a heart transplant, but Goswami and his team found that the tool may improve heart function capacity in these patients to the point that they may not need a transplant.

Much of Goswami’s approach to building AI tools involves looking at an outcome and working backward to determine how a patient or population arrives there. In the case of the Impella, the team wanted to build tools to predict whether patient heart function was likely to recover if the device was implanted, which could provide clinical decision support for providers determining if the device is a good fit for their patient.

Other projects the team at Mayo Clinic Florida is working on involve predicting whether patients with longstanding kidney disease can recover their kidney function with the Impella. At the time of writing, Goswami stated that six patients have had success with this approach and no longer needed transplants after recovering their kidney function.

These findings have led Goswami to advocate for reassessing transplant criteria in certain scenarios.

“Another thing that's important to understand is how we can use current technology in different ways to change the status quo of how people approach medicine,” he said. “What I'm specifically passionate about and really push here is [not always saying] ‘Hey, this patient is not qualified for X, Y, and Z.’ Because a lot of times, if you just make that blanket statement… you're really alienating a subsegment of the population that, based on traditional criteria from studies in the '80s or '90s, won't be included.”

AI and advanced devices like the Impella pump can create additional opportunities for these patients, who historically have had limited treatment options. Using predictive analytics to ensure that fewer patients progress to the point of needing a transplant is what Goswami hopes to achieve with these technologies. But, he underscored that such tools need to be studied and validated further to be adopted on a large scale.

A major part of the research and validation process requires access to vast amounts of high-quality data and teams of experienced data scientists to support model development, according to Byron Smith, PhD, a biostatistician involved in transplant research at Mayo Clinic in Minnesota.

Smith’s work focuses on understanding the clinical background for AI and deep learning (DL) applications in organ transplants. Currently, he and his team are working on improving outcomes following kidney transplants by prioritizing prolonged transplant kidney survival, patient survival, and patient well-being.

These efforts involve using AI to minimize the need for immunosuppressive medications and reduce side effects after transplant, such as nausea or tremors.

“[Following kidney transplant,] almost every patient has to take immunosuppressive medication,” Smith explained. “So, keeping the dosage lower to reduce the amount of side effects is of interest. We want these patients feeling good and healthy and happy with their kidney transplant.”

In pursuit of those three goals, Smith’s team is looking at factors related to patient death, kidney failure, and kidney rejection following transplant. Rejection status is typically determined via biopsy and typically quantified using a yes/no binary.

However, this doesn’t provide nuanced insights into the factors that influence rejection.

“A yes/no is not granular enough. What we really want to know is how much damage is done to the kidney,” Smith said. “And the pathology world has come up with schemes of saying, ‘This is an ordinal scale, so 0, 1, 2 or 3 for the amount of damage in certain parts of the kidney.’”

“Unfortunately, those scores amongst pathologists are not very reproducible. If you were to take it from pathologist to pathologist, they would give you different scores. So, what we wanted was an objective score that is continuous, and the best way to do that is to do it through the computer, through deep learning,” he continued.

Smith and his team are working on using DL to more effectively process biopsy images, compartmentalize damage, and identify features of interest within those damaged areas to inform scores that adequately capture the extent of a patient’s kidney damage.

These scores could then be used to build predictive models, improve medication development, find better endpoints for graft failure or patient death, and reduce the number of patients needed for clinical trials.

HURDLES TO ORGAN TRANSPLANT AI DEPLOYMENT

As with any large-scale health AI project, the models being developed by Mayo require significant amounts of high-quality, complete, and unbiased data. But, obtaining and managing that data is a major undertaking.

“So, there are a lot of different steps that we go through that involve a lot of different people in our lab to try to get the highest quality data that we can here at Mayo,” Smith noted, explaining that in his work to bolster kidney transplant outcomes, image biopsies must be scanned to produce high-quality images that can then be used for analysis.

From there, images are hand-reviewed to make sure that they are of high enough quality to be useful. Mayo also routinely takes biopsies at defined intervals for kidney patients, which helps ensure that patients are being monitored effectively to prevent issues before they arrive. This also provides a multitude of images for teams like Smith’s to use in their work.

The next step to ensuring data quality is removing observer biases, like gender and racial biases.

“We acknowledge that Mayo Clinic in Minnesota, Minnesota being a predominantly Caucasian state, we have a predominantly Caucasian population,” Smith explained. “However, our group especially tries to incorporate and collaborate with other people outside of the institution or at the other two [Mayo Clinic] sites to study different populations to make sure that we aren't completely biased or solely focused on that population.”

He noted that in collaborations with Mayo Clinic Florida and Mayo Clinic Arizona, which have higher Black, Hispanic, and Native American populations, researchers have discovered significant disparities and care barriers.

“These are some populations that were especially hit hard by things like COVID or have lower access to healthcare, and I think that those will be very important groups to study and understand better,” Smith underscored.

He further noted that developing AI tools that clinicians will be receptive to can be difficult, as providers have different preferences. However, both Smith and Goswami highlighted that advanced technologies like AI are designed only to augment the human touch in healthcare.

“One of the biggest things for people when I discuss AI with them is to understand that it's not going to replace a clinician. It never will,” Goswami stated. “I don't think people have to worry about a robot doctor in the future. But I think what people need to understand is that it's another tool in the armamentarium that we have to provide better care. And so right now, if I need to do a CT scan to say, ‘Hey, you have lung cancer,’ or I could use an AI algorithm in the clinic with a drop of blood that takes 30 seconds to run, that's what I'm going to try to do to be able to optimize the care, right? Because the faster we can diagnose something, the more we can focus our treatment on it to cure you.”