Patients with rare, aggressive diseases can be a challenging opportunity for care providers, doctors, and treatment specialists. Lower incidence rates and difficult treatment processes can lead to more expensive care and poorer prognosis rates, leaving those diagnosed feeling like the odds are stacked against them from the start. Is AI transformative? Can it help with rare disease care?
However, artificial intelligence (AI) may just be changing these odds for the better. From quickly sorting through patient information to cut down on administrative costs, to diagnosing cancers faster than leading oncologists are able to, AI certainly seems like the best path forward. Understanding exactly how a computer will revolutionize rare disease care may seem like a daunting task, but it’s worth undertaking.
How Does AI Work?
The field of AI may seem confusing or even unfathomable, and articles that use it as a buzzword without explaining exactly how the technology works don’t help the situation. It is not one specific program or machine, but rather a method for how algorithms can be built to mirror human thinking and intelligence.
Artificial Intelligence as we know it today first appeared in a program called The Logic Theorist, which was designed to mimic human problem solving and learning skills. The program was presented at the 1956 Dartmouth Summer Research Project on Artificial Intelligence conference, which served as the beginning of modern AI advancements.
Prior to the 1950s, artificial intelligence was certainly on the horizon, most notably in the form of British mathematician and logician Alan Turing’s 1950 paper Computing Machinery and Intelligence. However, Turing’s machines lacked the ability to store their commands and extrapolate further data from them, instead merely executing the human input commands.
Turing’s work, which articulates the theory that human problem-solving ability is based on available information and reasoning from past experiences, illustrated the basic concept of a neural network. Neural networks in early AI were designed with the human brain as a muse, programming the machines to connect and store given data, and then later use the saved data as a reference to help solve new problems.
Photo by Pexelson
These problem-solving networks allow machines to learn from previous experience and become “smarter” as their catalog of stored data grows. This is the field of “machine learning”, which is a sort of category within the larger realm of AI. For machines to learn, they skim through data and use pre-programmed algorithms to make informed decisions. Machine learning algorithms are ever-present in daily life, using AI to suggest our next song on Spotify or target ads on the internet.
Deep learning is the next level of AI, which has driven the boom of the last decade, and appears to have the most potential for use in disease care. Lighter forms of machine learning require human input to confirm decisions and learning, but deep learning allows the computer to act entirely on its own.
Machines equipped for deep learning have tremendous potential in the medical field, much like an infant’s flexible and hungry mind. For successful deep learning models, there are a few important requirements. They must have access to huge amounts of data and immense computing power, so currently only large research firms and universities are in the playing field of deep learning AI.
Where Does AI Have the Most Potential?
Machine and deep learning are both subsets of the overarching field of AI development and can be used in all sectors of human innovation. Financial and stock trading programs, consumer forecasting, leisure planning, and manufacturing are just a few of the many industries that can benefit from this massive potential.
Healthcare seems to be utilizing AI in its most evolved form. Any advances made in the medical field to harness this power could have far-reaching applications, too. Medical research is one of the biggest data-driven fields today, and the immense number of studies and a large volume of cataloged patient data makes it a natural fit to further develop deep learning machines to reach their highest potential.
However, a Google search for “AI in Healthcare” may be more misleading than it is accurate. Separating the exciting new discoveries from exaggerated assertions of medical breakthroughs is difficult, but worthwhile if the power of AI is ever to be properly harnessed and maximized for patient care.
AI in Healthcare Administration
In 2015, the healthcare industry worldwide was estimated at a little over $7 trillion in annual worldwide spending, and according to Deloitte, that number could potentially reach $8.734 trillion by 2020. The Center for Medicare and Medicaid Services also reported in 2016 that the average American spent $10,345 on healthcare annually, and a cancer diagnosis can inflate that cost considerably.
Using smart machines to automate administrative healthcare tasks could result in savings of up to $18 billion, reported Forbes. Other healthcare savings amounting to $20 billion using virtual nursing assistants and $40 billion with robot-assisted surgery demonstrate the extreme force of change that integrating machines into healthcare could have.
Nuance Communications, Inc., debuted their Dragon Medical Virtual Assistant in early 2018 to help streamline clinical workflow and cut down on administrative crowding. This virtual assistant was designed specifically for healthcare providers with features specific to enhance security and organization. VentureBeat wrote that systems like Dragon Medical “offer real-time clinical documentation guidance that helps providers ensure their patients receive an accurate clinical history and consistent recommendations.”
Cancer in the Clouds
Data organization and access is perhaps the biggest hurdle oncologists and treatment development professionals have to navigate when treating rare diseases. Incredible amounts of data on patient scans, cancer studies, and medical research exist with no single centralized system making it available for researchers around the world.
The Cancer Genome Project (TCGP), which began in 2005 and is supervised by the National Cancer Institute (NCI), catalogs cancer variations and the genetic mutations that cause each case. TCGP is one of the most concentrated efforts to analyze and compile the huge amounts of available data in our cancer-fighting history.
Recognizing that TCGP perhaps gathered more information than a human could sift through in an entire lifetime, the NCI launched a competition in 2013 to design a cloud computing system to store and organize all of this collected information. The winners were the Broad Institute with UC Berkeley and UC Santa Cruz with $7 million, Google partnered with the Institute for Systems Biology to receive $6.5 million, and Seven Bridges Genomics received $6 million.
These efforts to organize cancer data are part of an exciting push to make sense of all the available data to find the keys to defeating cancer. Although these systems are up and running, they’re not universally used yet and aren’t integrated with some of the more exciting detection and diagnosis tools that use AI today. Aligning these two huge projects could finally turn the tide against cancer.
Disease Detection + Diagnosis
If you’ve seen headlines about AI’s potential to cure cancer once and for all, they probably cited exciting rates of disease detection and diagnoses by AI-based machines. These systems, when equipped with deep learning algorithms to parse through patient scans and detect minute abnormalities based on previous patient data, certainly may be a huge step forward for cancer.
Computer engineers at the University of Central Florida (UCF) programmed a computer to detect tiny specks of lung cancer from CT scans, oftentimes better than human radiologists could. In fact, the computer was accurate 35 percent more than the humans were. These UCF engineers aren’t the only ones showing these kinds of incredible results. Researchers all over the world are applying advanced algorithms to detect skin, prostate, and breast cancers more quickly than doctors are able.
Even rarer forms of cancer, like asbestos-caused mesothelioma cancer, are getting a piece of the AI game. Canon Medical Research Europe has jumped into the deep end of the pool, gaining £140,000 to develop an AI-powered assessment tool for malignant pleural mesothelioma (MPM). Mesothelioma hits the trifecta of terrible qualities to have in a disease: it’s rare, aggressive, and moves unpredictably. Life expectancy for MPM is typically less than two years after diagnosis, so earlier diagnoses could extend life expectancy considerably and greatly improve patients’ quality of life.
Can AI Tailor Treatments?
After detecting and diagnosing cancer, the next step is to create a treatment plan that will act as quickly and accurately as possible, while doing the least amount of damage to a patient’s body. Researchers at the University of Toronto developed a system using AI to analyze historical treatment data and develop radiation treatment plans for new cases.
This team’s AI was able to produce comparable treatment plans to those developed by humans. Radiation therapy is an extremely taxing treatment for patients and requires extreme care and specificity to be successful without irradiating healthy parts of the body. Personalized cancer treatment developed more quickly saves lives.
However, AI-based disease treatment is definitely in more infantile stages than detection and diagnosis. In fact, some teams’ efforts to use this technology have turned out less than ideal.
Artificial Intelligence and machine learning (AI/ML) may be at the “peak of inflated expectations,” as New England Journal of Medicine authors Jonathan Chen and Steven Asch put it. Here lies Watson, IBM’s proprietary AI wonder child that was supposed to “define the field of cognitive computing.”
When used to recommend treatment plans for cancer patients, oncologists complained that Watson made “unsafe and incorrect treatment recommendations,” going so far as to tell doctors to use blood-thinning medication on a hemophiliac patient, which would cause extensive blood loss. These recommendations, though made for hypothetical patients, represent an alarming inability to quantify basic medical needs.
IBM Spokeswoman Christine Douglas admitted Watson’s shortcomings but continued to hedge that the system’s potential is still being developed. She said, “the opportunity for A.I. in health care is still nascent, but we are proud to be pioneers in this arena.”
Indeed, perhaps this failure is less a result of Watson’s coding, but more reliant on the fact that the system was trained on fabricated medical records. Unless research teams work directly with hospitals or projects that provide real patient data, getting their hands on real-world medical records is very difficult due to patient confidentiality protections and HIPAA laws.
How to Regulate AI for Healthcare
Watson’s unsafe recommendations, though recognized by doctors and not applied to an actual patient, opened a can of worms for medical regulatory bodies. If AI systems become standard practice that affects real patients, who is liable if they fail?
At a health and data conference in April 2018, U.S. Food and Drug Administration (FDA) Commissioner Scott Gottlieb gave a vote of confidence for AI and indicated that the FDA is working on this newfound quagmire of medical device regulation.
“AI holds enormous promise for the future of medicine,” Gottlieb said, “and we’re actively developing a new regulatory framework to promote innovation in this space and support the use of AI-based technologies. So, as we apply for our Pre-Cert program – where we focus on a firm’s underlying quality – we’ll account for one of the greatest benefits of machine learning – that it can continue to learn and improve as it is used.”
Photo by Pexelson
To radiologists, it may seem like a doomsday sentence to encounter a computer that can process information in the blink of an eye that would take the human brain years to comb through. However, the possibility for radiology and oncology to see positive growth and change with the aid of computers is vast.
A comfort to radiologists worried about job security is the pure point of liability. Misdiagnosis and missed diagnoses can be huge burdens on a hospital. Even if AI becomes common in treatment facilities across the globe, hospitals still want a human set of eyes to double and triple check to eliminate any margin of error.
The Medical Futurist Blog put it succinctly; “radiologists who use AI will replace those who don’t.” The real threat to radiology jobs are not machines, but rather the wave of radiologists who are comfortable deploying this technology. Other humans are the competition, not computers, so adapting and embracing new technological advances is the best way to stay relevant in the changing world of radiology and healthcare.
Using the Cancer Genome Project’s new cloud systems to bring real patient data to researchers could avoid and eliminate pitfalls like those experienced by IBM’s Watson. New regulations could allow for the careful implementation of AI into medical practices, and a growing realization the AI is the future could bring more funding to niche tools, like Canon Medical’s mesothelioma detector.
The machines may not be coming for medical jobs, but hopefully, they’re coming to hospitals and treatment centers to improve patient health. Once all of these disparate advances in AI and healthcare IT are centralized, modern medicine could finally be on the verge of having a disease-buster, so to speak, in its hands.