Artificial Intelligence in Medicine
Contributor Vivian Cheng
Science fiction novels have thoroughly explored worlds that coexist or wage war with artificial intelligence (AI). From Blade Runner’s anxiety-inducing replicant-filled world to Neuromancer’s technologically savvy society, AI has been characterized as a world problem solver to a trouble-maker that introduces a myriad of ethical concerns.
Interestingly, Canada is a world leader in artificial intelligence. Geoffrey Hinton, the Google Engineering Fellow and University of Toronto cognitive psychologist and computer scientist, is considered the godfather of deep learning because he has assembled neural nets in a way that mimics the human brain.
Contentious debate aside, AI provides vast opportunities through its ability to analyze complex health data in medicine and assist with hands-on procedures. With large untapped sources of data in healthcare, AI has the potential to examine a large array of cases and transform the healthcare system. With that being said, transformation doesn’t often come in ways that we expect.
How does AI affect the medical field? What areas of medicine have been affected by automation already?
AI projects come in different sizes, shapes and forms, but all have the common goal of improving patient outcomes. AI in medicine falls into two main categories: virtual and physical. Virtual refers to informatics approaches, such as information management to control of health management systems. Health management systems include electronic health records and systems that provide active guidance of physicians in their treatment decisions. In contrast, the physical form generally consists of more hands-on approaches, such as machine assistance in surgery.
Some of the most ambitious versions of diagnostic machine-learning algorithms seek to read patient’s medical records and compare the information against an encyclopedic database of medical conditions from textbooks, journals and medical databases.
Human error and long wait times can all drastically impact individuals’ health. Take for example stroke patients. Doctors say that diagnosing a stroke once the brain is dead and gray is easy, but much harder before too much brain has died. On a diagnostic scan, blurring at the borders between the anatomical structures may show up, but that shadow usually appears on the scan several hours, or even days, after the stroke — when intervention ceases to be helpful. Before this stage, the scan shows only the premonitions of a stroke.
Diagnosing strokes too late can result in life-changing health problems, like paralysis.
In the U.S., a firm called Viz.ai uses machine learning to analyze brain scans in order to categorize stroke patients based on the urgency of their situation.
When strokes occur, brain cells are lost every minute the clot remains. However, ordering effective therapies like clot-busting drugs and thrombectomies rarely occurs because too much of a patient’s brain has died by the time the stroke is diagnosed. Viz.ai attempts to mitigate this issue by sending brain scans directly to specialists.
AI is not limited to stroke diagnostics, however. Many fields in medicine, like emergency medicine, oncology, ophthalmology, cardiology and more are trying to capitalize on AI, by improving diagnostic and treatment tools. To read more about current innovations, check out Artificial Intelligence in Medicine.
At Stanford University, a software was trained to classify skin lesions, and it was found to identify keratinocyte carcinoma (the most common type of skin cancer) and malignant melanoma as well as the dermatologists. But one AI-skin-cancer-detection system can do even better. The Annals of Oncology reported that an AI system, when pitted against 58 dermatologists, found 95% of skin cancers compared to 86.6% found by humans. As well, it misdiagnosed fewer benign moles as malignancies.
Similar technological developments are being made for breast cancer, eye disease (i.e. glaucoma and diabetic retinopathy), heart disease and cardiac arrhythmias.
Diagnostics only show one side of AI’s virtual assistance capacities. Decision support systems, laboratory information systems and robotic surgical systems also exist. These systems guide diagnostic/treatment assessments, take note of infection rates and assist with surgery, respectively.
Here in Ontario, at St. Michael’s Hospital in Toronto, Dr. Frank Rudzicz and his team of researchers investigate machine learning and natural language processing in hopes of producing software that helps individuals with disabilities communicate.
In the future, diagnostic devices may also become portable in your own smartphone. Researchers at Stanford University, along with a firm making portable ECGs is helping Apple see whether arrhythmias can be detected in the heart-rate data picked up by its smart watches.
What are the benefits? What are some of the issues in AI that scientists are trying to mitigate? What issues must be resolved before the technology enters the market?
As AI improves, its innovations may change the healthcare field, by taking the drudgery and error out of diagnosis. Although humans and computers may have similar diagnosis accuracy rates, computers may arrive to conclusions much more quickly. This feat may, in part, tackle the problem of our growing senior population and patient volume.
AI may also push the future towards precision medicine (Check out Ramtin Hakimjavadi’s article about it for more information) because they are able to be more specific than humans, by categorizing illness by severity for example. Our current approach to treatment is based on ‘the needs of the statistical average person’. However, personalized treatment can become more feasible once machines can collect and analyze data more quickly.
As algorithms continue analyzing new cases, they become increasingly more accurate. However, humans must ensure that machines learn from representative datasets to avoid biases that may have detrimental effects on human health outcomes. Machines are only as good as the datasets they learn from. Datasets used to develop these learning programs can also be costly for researchers to procure. In Ontario, the Vector Institute obtains their dataset of Ontarians through Collaboration with the Institute for Clinical and Evaluative Sciences.
There is also the “black-box problem” that’s endemic to deep learning. Scientists do not know the exact mechanisms at which AI arrive to their conclusions. Machines just arrive at answers.
Before AI enters the market, an AI model also must be developed and tested through clinical trials, to provide guidance for building other AI machines. Regulatory bodies must also approve these technologies.
Finally, health-care institutions must purchase AI machines. Given the exorbitant costs of the healthcare system, widespread adoption of AI machines is unlikely to come anytime soon, at least, not all at once. Single-use or disposable surgical supplies (i.e. sutures, scalpels and sterile drapes) can amount to thousands of dollars per case and result in the addition of millions to hospital expenditures. Adding new technologies will likely overwhelm an already overburdened system.
What are the legal concerns?
Although machines can master the facts and interpret the facts, they cannot tell us why they come up with certain answers. If we’re to create or modify policies and frameworks that govern and regulate machine learning, healthcare and etc. in the future, we should be able to understand the mechanisms and risks of machine learning.
We need to ask real questions about how negligence, liability, intellectual property and other regulatory laws will change when AI becomes more widespread.
As well, we need to address privacy concerns as AI collects more sensitive patient information. Currently, information is regulated by statutes like the Personal Information Protection and Electronic Documents Act (PIPEDA). You can read more about the legal aspects of AI here.
How does AI affect human employment in the short run and long run?
In a similar vein, AI will not make human experts redundant for the time being. Even as AI gains the ability to complete more tasks, these tasks will require careful monitoring.
Currently, health care provides one of the largest sources of jobs, and is predicted to be the largest contributor to job growth in the next decade.
In Canada, the public will experience a worsening doctor shortage, requiring more, not fewer, providers, so technology may actually alleviate the burden on health-care providers, not replace them.
How will AI change the doctor-patient relationship?
The doctor-patient relationship could change in many different ways, and it’s difficult to make accurate predictions on technology that’s still far from implementation. Drastic changes may occur or they may not occur at all.
On the one hand (and perhaps the more obvious answer), health professionals’ may be allotted more time to do tasks seen as more valuable because AI reduces diagnostic times. Perhaps their care may become more humanistic and patient-oriented.
On the other hand, because AI requires careful monitoring, doctors may still spend approximately the same amount of time with their patients. As part of their informant duties, doctors may also have additional responsibilities in explaining AI’s mechanisms, risks, benefits and etc. Doctors may also have to alleviate their patient’s fears regarding this technology.
How does AI affect residency programs?
The nature of medical education will likely change as well. Doctors learn to make diagnoses by first studying medicine, collecting facts about patients’ conditions, generating a list of potential causes, then eliminating certain hypotheses. Doctors consider the likelihood of certain diseases as well as the patient’s history, risks and exposures to come up with assessments. We don’t know how integrating machines into the practice will change the scope of their practice.
As many of you gear up for your medical school applications, you may be wondering how AI will affect residency training. Many teaching hospitals in the U.S. have not yet implemented education on AI for its young doctors. The statistics are unknown in Canada.
However, in Canada, some forms of AI are already in hospitals — namely, surgical robots.
One paper from earlier this year examines the barriers residents had when it came to learning and interacting with surgical robots in their training.
Traditionally, residents learn from surgeons by providing hands-on assistance in the patient and they also learn when they get the chance to operate.
However, when it came to robot-assisted surgery, young surgeons were stuck at bedsides or sitting in a second trainee console watching the operation. In both cases, they had decreased chances to operate for robot surgeries.
When residents did get a chance to operate, their teaching opportunities were of lower quality because because doctors would give frequent criticisms and take control of the robots. Many residents had trouble overcoming these barriers. The ones that were successful focused on robotic surgery in the midst of their medical school, or practiced via external stimulators and watched recorded surgeries on YouTube — characteristics that define “shadow learning.” However, none of these strategies were openly discussed. The more successful residents became hyper-specialized in robotic surgeries, at the expense of more general surgical skills.
Despite these issues, most doctors at these teaching hospitals became professionally empowered to perform robotic surgeries when they finished their residency.
While this example may provide relevant insights into the field of surgery, they can be extended to other medical fields with AI. Issues may emerge such as increased prevalence of shadow learning and a hyper-specialized minority.
One thing’s for certain: as AI becomes more popular in healthcare settings, residency programs should adjust to better accommodate AI education in their training programs.
Despite the fictional content of Westworld and Ex-Machina, AI is emerging in many spheres of our lives — and medicine is no different. As more healthcare professionals and leading academics talk about the future of healthcare, AI is sure to come up again and again. While AI has the potential to bolster productivity, slash wait-times, and solve a whole slew of healthcare issues, we must remember that AI is still in its infancy, and much more research still needs to be done before AI is readily adopted in our clinical settings.
While it might be fun to look into our crystal ball to see what the future holds, for now, we shouldn’t get too caught up in future predictions about AI.