​ Health Information Technology

Artificial Intelligence

A new medical colleague

BY SISI MA, PHD, CHRISTOPHER TIGNANELLI, MD, AND DEMETRI YANNOPOULOS, MD

or centuries, the practice of medicine has relied on human intelligence to diagnose disease and prescribe effective therapies and curative strategies. Decade after decade, and through years of advances in research, we continue to expand the level of human understanding of medicine. Unfortunately, even as physicians with the best intentions, the adage to err is human applies. It’s a difficult truth, but one that is changing in unexpected ways.

That change involves using an additional type of intelligence — one made by humans to be smarter than humans, otherwise known as artificial intelligence. When Stanford professor and computer scientist, John McCarthy, first coined the term AI in 1956, it’s hard to believe that even he understood the implications that it would have decades later. Today, AI is transforming industries around the world, and the field of medicine is beginning to employ AI as a critical tool in research and patient care. From making more accurate diagnoses and increasing workflow efficiency, to determining risk for disease and identifying the most effective treatment options, AI simply brings more precision and speed to a field that requires both, oftentimes, in extremely high-risk and life-changing situations.

To provide better health care to all patients, it’s imperative that physicians embrace AI. AI can and will impact every medical specialty and the University of Minnesota Medical School is a global leader in driving this innovation. What follows are examples from three different fields summarized by the project leaders.Assessing the method of care required

The world’s first AI-controlled CPR system

Despite the progression of medicine elsewhere, the act of CPR remains largely unchanged, even after 60 years as the first-line practice for those experiencing cardiac arrest. But, the numbers prove a change is necessary — after 20 to 25 minutes of CPR, the chances of surviving are in the single digits. That’s because it provides only 10 to 20% of blood flow initially, then deteriorates over time.

​AI can and will impact every medical specialty.

My team, including researchers from both the Department of Medicine’s Center for Resuscitation Medicine at the U of M Medical School and the Georgia Institute of Technology, led a study that has helped develop the world’s first AI-controlled CPR closed-loop system. Our early results show that the system performs better than both a licensed physician and the LUCAS machine — the current standard of care — during a cardiac arrest. For 20 years, I have studied resuscitation medicine and began recognizing a need to change the “one-size-fits-all” approach to CPR. But I’ll admit, when I first heard about AI, I was skeptical. Over time, though, the data began changing my point-of-view.

In our study, we trained our algorithm in pre-clinical studies, ultimately noticing superior coronary perfusion compared to other CPR methods. For example, at the 15-minute mark, as the CPR provided by the physician and the LUCAS machine trended downward, the AI machine-controlled CPR trended upward and stabilized at higher rates of coronary perfusion pressure. In other words, this system outperformed physicians who could manipulate the same patterns of CPR while knowing the same information that the computer does. Why? Because the system has the ability to calculate and assess data a million times per second, learn from what it does and then predict the future to optimize the simple concept of chest compressions. A human brain simply cannot do this.

We’re not finished yet. These early studies did not take into account compression rate, depth and timing. Instead, we focused on the depth of chest compressions and decompressions. Over the next three years, we will factor in these additional components of CPR into the algorithm to continue to optimize the system. And, when you really think about that — about just how complicated this algorithm will become — it will likely be almost impossible for a human to comprehend. Yet, it will save thousands of lives.

Diagnosing COVID-19 faster

The lives of those lost during COVID-19 can be partly attributed to testing limitations. AI brings new speed to medicine, and when it comes to testing, diagnosing, and treating patients with COVID-19, speed is critical. These testing limitations, caused by a variety of external factors, prevent many from knowing whether or not they carry the virus, leading to increased spread and, for some, worse health outcomes.

​

A team of mine, as part of the U of M Critical Care Outcomes & Research Effort, began leveraging AI and health informatics, before the pandemic, to study ways to improve health care outcomes. As we neared launching a project that leveraged AI for traumatic injury detection, Minnesota hospitals began admitting their first COVID-19 patients — we knew we needed to pivot. We wondered if our AI research could provide a solution to support the testing demand.

In a handful of months, we leveraged our AI infrastructure and research team — a collaboration of U of M experts in computer science, radiology, surgery critical care and health informatics — to change our traumatic injury detection system to instead diagnose possible cases of COVID-19 using chest X-rays. With the help of a research grant awarded from Microsoft Azure’s AI for Health, our team trained the algorithm on approximately 40,000 X-rays of patients who did not have COVID-19 and 5,000 X-rays of patients who did, thanks to available de-identified patient records through our M Health Fairview partnership.

Once our team validated the algorithm, we partnered with M Health Fairview leaders and Epic to build an infrastructure around the algorithm, designing it to seamlessly and immediately translate the algorithm’s findings into the medical records software. That way, when a patient arrives in the emergency department with suspected COVID-19 symptoms and clinicians order a chest X-ray as part of standard protocol, the algorithm will automatically evaluate the X-ray, recognizing patterns associated with COVID-19, and — within seconds — notify care teams through Epic that the patient likely has the virus. This algorithm is currently live within the M Health Fairview healthcare system for investigational use.

We hope to receive FDA approval soon, and when that happens, M Health Fairview has agreed to make this tool free of charge to more than 450 hospital systems around the world that use Epic. Each hospital system will be able to download and install the algorithm in as little as 10 days. This tool, fueled by AI, will prevent unintentional exposure to COVID-19 for staff and other patients in the emergency department, and possibly, help supplement diagnostic testing that, still today, faces supply chain issues and slow turnaround times across the country. Even health systems in low resource areas with high infection rates and less access to testing could use the tool, helping us fight COVID-19 in underserved communities. Overcoming COVID-19 requires all of us to work together, and AI is, and will continue, playing a critical role in that.

​In other words, this system outperformed physicians...

Better therapies for adolescent depression

We know AI can help diagnose disease, and we know it can improve out-dated medical practices. But, can it also support a physician’s decision-making on the most appropriate therapies for each patient? Yes.

My collaborators and I at the U of M Medical School recently developed an AI, machine-learning tool that identifies the best treatment option for adolescents with depression. A significant public health problem that continues to rise, growing from 8.7% in 2005 to 11.3% in 2014, adolescent depression increases suicide risk and, without proper treatment, can greatly interfere with academic, social, emotional, and neurobiological functioning that can lead to a lifetime of physical and mental health impairments.

Yet, despite the availability of several evidence-based treatments, including psychotherapies, medications, and their combination, the overall response rates are alarmingly low — about 30 to 50% of adolescents do not respond to therapies. Ineffective treatment discourages many patients from pursuing further treatment and unnecessarily exposes patients to medication side effects, undue expenses on health care services, and lost work and/or school time. This low response rate can be attributed to the nature of depression. Because it is a heterogeneous disorder with multiple etiologies and symptom profiles, to effectively treat each patient, special attention must be given to the individual characteristics that influence each patient’s depression — a job made for AI.

Our algorithm optimally matches treatments to patients’ individual characteristics, needs, and circumstances, including the severity of depression, patients’ medical history, comorbid diagnosis, family dynamics, social functioning and more. We leveraged the only large clinical trial with an extensive baseline assessment battery, called “The Treatment of Adolescents with Depression Study (TADS),” that compares the three primary treatments for adolescent depression, including cognitive behavioral therapy (CBT), fluoxetine, and a combination of both. The resulting model successfully identified subgroups of patients that responded preferentially to specific types of treatment. Based on the CDRS-r scale, which quantifies depression severity, a subset of patients achieved, on average, a 16.9 point benefit to fluoxetine compared to CBT and another subgroup of patients that achieved an average benefit of up to 19 points from a combination compared to CBT.

Our next step is to validate and apply these models in the clinical setting by finding a way to integrate the algorithm into the patients’ electronic medical records, so that the process is automated and user-friendly. Providers would gather the needed information from the patient as part of regular care and input it into the electronic medical records. The AI will take the relevant information, predict treatment response and give recommendations about different treatment options. Clinicians and patients could then examine the model predictions and recommendations together to select the preferred treatment.

​

We think this AI model provides a critical starting point for the personalized treatment planning of adolescent depression, but we anticipate that this methodology could be used to examine other clinical trials and make discoveries about differential treatment responses for other diseases and treatments as well.

A new symbiosis

Embracing AI does not mean we will be replaced as physicians. Although we agree at the U of M Medical School that a much-needed symbiosis with AI will expand our ability to provide for patients in ways like never before, we need to complement AI with humanism — our common sense, optimism, and compassion for the patient in front of us. That is something computers will never do, but with our wisdom and instruction, it can help solve some of the most pressing health issues of today and in the future — if only we are open to it

Sisi Ma, PhD, is an assistant professor in the U of M Medical School’s Department of Medicine and Office of Academic and Clinical Affairs’ Institute for Health Informatics. Her primary research interest is the application of statistical modeling, machine learning, and causal analysis methods in the field of biology and medicine. She worked on the adolescent depression project.

Christopher Tignanelli, MD, is an assistant professor in the Department of Surgery at the U of M Medical School. He is also a general surgeon with M Health Fairview, specializing in trauma, ECMO, emergency general surgery, and has research interests in surgical outcomes and patient safety. He worked on the COVID project.

Demetri Yannopoulos, MD, is a professor in the Department of Medicine at the U of M Medical School. He is also an interventional cardiologist with M Health Fairview, specializing in emergent cardiac care, coronary-artery disease, and congenital and peripheral intervention, and has research interests in cardiopulmonary resuscitation, hypothermia, and myocardial salvage during acute coronary syndromes. He worked on the CPR project.