Machines with Brains
There’s a very simple calculus involved in nearly all aspects of health care: for every X patients, we need Y doctors. But, in the US, when it comes to cancer treatment, it’s becoming abundantly clear that the math no longer adds up.
Experts predict that by 2020 the demand for cancer treatment in the US will far exceed the supply of oncologists who can provide it. While the workforce’s capacity to treat patients will grow by 14%, that’s nowhere near the 48% growth needed to support an aging population and a rising number of cancer survivors—whose lives have been saved by improved treatment options, but who require continued care or whose cancer may recur, according to the American Society of Clinical Oncology. Other fields related to cancer, like radiology, are also expecting increased demand.
Elsewhere in the world, the situation is already dire. Just north of the US in Ontario, Canada, just 185 radiation oncologists treat the entire province’s cancer patients; there were 72,000 new cases diagnosed in the province last year, among a population of nearly 14 million. (For comparison, New York state has more than 3,200 radiologists to serve its 19.8 million residents.) The city of Durban in South Africa has lost its last publicly-funded radiation oncologist, leaving the uninsured (and 84% of South Africans do not have insurance) among the 11 million who live in the city and surrounding province without access to care.
Cancer doctors are already overburdened, and it will get worse. In addition, thanks to rapid advances in techniques like genetic sequencing and data storage capabilities, we now have access to an unprecedented amount of information about cancer and the patients they afflict—but we don’t have the manpower to make use of all these data in a way that would lead to the sorts of personalized treatments that the genetic revolution promised.
There are two obvious solutions: Train more medical professionals, or make existing workers more efficient and experienced beyond their years. Technologists are focused on the latter; changing the whole medical school system might be beyond even the most ambitious of Silicon Valley tech disruptors, so instead they are working to develop artificial intelligence to help doctors make better decisions faster, and to parse the massive amounts of genetic data currently sitting underused on servers and hard drives around the world.
Artificial intelligence is really good at identifying patterns. When an image-recognition algorithm sees thousands of pictures of horses, it begins to learn the patterns of pixels that typically represent a horse. If the algorithm is shown images of horses from different angles, in different lighting, and against different backgrounds, it will more fully understand the idea of what a horse looks like. Moreover, these algorithms can be made tremendously sensitive to small patterns, like what differentiates an Arabian or Clydesdale—or like the minute differences between two medical scans.
Doctors can spend hours analyzing just one scan of a patient’s brain, looking for subtle patterns that could indicate a tumor. When artificial intelligence algorithms are trained to find those minute patterns, they can help determine whether there’s a tumor on a CT scan faster than a human, or look at a smartphone image of a skin lesion and tell with the confidence, speed, and accuracy of a trained dermatologist whether it requires a biopsy.
Caution is always called for when it comes to introducing any novel concept in the critical field of medicine, but preliminary tests have shown these automated systems to be as effective as highly trained physicians at some tasks, like reading X-rays and retinal scans, and predicting whether a patient’s cancer will relapse. Researchers from Stanford University pitted an AI they built to diagnose images of skin lesions as either malignant or benign against 21 board-certified dermatologists. The algorithms beat the humans nearly every time, research published in January showed. The goal is not to replace doctors, but to give an instant second opinion without sending a patient to a specialist—or be used by someone without a medical degree to provide basic diagnostic services in remote or underserved communities.
Research to date and early products have focused on maximizing accuracy by building one algorithm that looks for a specific disease on a specific kind of scan. Clearview Diagnostics, for example, is currently building AI-powered software that radiologists could consult as a sort of second opinion on whether a growth is more likely malignant or benign, and whether the probability of malignancy is high enough to warrant further tests.
As these various specialized algorithms proliferate, in the not-distant future, they may form a digital toolbelt that give general practitioners instant second opinions—for example, they might see a suspicious-looking mole and run it through the algorithm to see if it suggests a high risk for melanoma. If the algorithm has high confidence it’s not malignant, patients could be saved from an unnecessary biopsy. Or, an oncologist could use an AI service to parse a patient’s medical history and then predict how that patient would respond to a certain treatment. In areas where medical coverage is less available, faster and more precise decision-making means higher productivity for doctors. These algorithms would become part of a tapestry of hundreds of supporting technologies that will shape the experience of going to the doctor in the 21st century, much like the stethoscope and otoscope did in the 20th.
Unlike those pieces of hardware, algorithms are looking to supplant parts of a doctor’s expertise—which means the implementation of these algorithms is an uphill battle. It’s been shown time and time again that humans simply don’t trust an algorithm’s decision as much as a human’s, a phenomena known as algorithm aversion. Proposed FDA rules (pdf) for medical software cite accuracy as a first priority, and would mandate the accuracy and reliability of the software at least fall in line with the current medical standards.
Rick Mammone, CEO of Clearview, says the algorithms also can act as a safety net, catching abnormalities that a human might have missed. “If they’re distracted, if they’re tired, if they’re reading too many [mammograms], if they’re not trained for as many years as someone else, all these variables become irrelevant because they have a baseline to look at,” Mammone says.
Speed is also factor: In situations where tumors are growing rapidly, a few days shaved off the diagnostic timeline could be a matter of life or death. AI has the potential to speed up the diagnosis-to-treatment in a number of ways. For one, getting a second opinion on a laptop in the doctor’s office removes the effort of seeking out another doctor and all the moving of medical records and human bodies the process usually entails. In addition, algorithms have the ability to rapidly speed up the treatment planning phase that comes after the initial diagnosis.
Planning any cancer treatment, from targeted radiation to surgery to chemotherapy, begins a complex conflict between the doctors and cancer, the dynamics of which can change almost daily, says Stanford’s Jeff Shrager.
“You make a move, and cancer makes a move. It’s difficult to predict what cancer is going to do, once you do something,” Shrager says. “Treatment planning has a fairly short horizon.”
The sequence of treatment depends on the location and type of cancer. Radiation is often used in a targeted way. In order to be to be effective, radiation oncologists need to know the exact dimensions and placement of the tumor, leading to a task called segmentation.
Radiation oncologists typically take hours on segmentation, drawing lines on a 3D scan to indicate the boundaries between a tumor and healthy surrounding tissue. Artificial intelligence researchers have been working on a potentially applicable technique that coincidentally goes by the same name as this line-drawing practice: “segmentation.” It’s a key concept in AI; for AI to determine what an object is, it has to know its shape, and where it begins and ends.
Large technology companies including Google DeepMind have begun to tackle this task; DeepMind and others have been able to cut what is a typically four-hour process, in delicate areas like the head and neck, down to a matter of minutes.
AI can also help mitigate the health-care challenges that come from the fact that cancer isn’t one specific illness, but an umbrella term for a collection of diseases defined by the rampant and uncontrolled replication of cells. Tumorous cancers of different organs behave differently; there are cancers of body parts that aren’t organs (like leukemia, a type of blood cancer) that behave differently than those that cause tumors; and as scientists learn more about genetics and biomarkers, they are beginning to understand there are differences within the traditional cancer groupings—breast cancer, for example, is now divided into four different molecular subtypes, each acting differently and entailing a different treatment approach.
Both the medical field and outside tech companies have realized the economic potential of a system that could draw on the complex history of cancer care and recommend a treatment that’s been successful for similar patients in the past. The goal: To simulate the expertise of a doctor who has seen thousands of patients—and remembers the details of every single one—and read every medical text available—and never forgets a word.
To patients, this means their future will likely be decided by a machine. While a doctor might deliver the news and explain how the disease and how treatment will work, the actual diagnostic and planning work of the medical professional will likely transition to making sure the machine hasn’t made an error. Some in the field believe AI will also be able to monitor treatment progress, with a nurse or nurse practitioner guiding patients through the process, freeing oncologists to do the more specialized work they’ve trained for. But it’s worth noting that this technology hinges on human knowledge: flaws in data or in the algorithms can cause the same errors a human doctor might make.
IBM has been one of the most vocal technology companies pursuing the goal of automated treatment, through its “Watson”-branded constellation of tools. When a specific patient’s medical data are loaded into the Watson system, custom reports can be generated with different options for treatment, including experimental trials.
Oncologists themselves have recognized the potential applying machine learning to parsing and analyzing big medical data. The American Society of Clinical Oncology (ASCO) has created its own platform called CancerLinQ, which now has more than 1.5 million de-identified patient records from almost 90 US-based contributing oncology practices and academic institutions. Once the database collects enough information to make reliable decisions (it could take years, ASCO says), oncologists will be able to enter in their patient’s health data, and get customized treatment suggestions.
That’s a major change from current best practices. “If a patient comes into the room and sits in front of me, say it’s a 70-year-old woman with lung cancer, I’m basing my information on groups of data from clinical research trials plus my own personal experience and the experience of my colleagues by attending tumor boards,” says Robin Zon, an oncologist and ASCO committee chair. “But it’s not necessarily precise to her situation.”
Zon says additional data like molecular information of the tumor or granular medical history analyzed by the computer could lead to a treatment that’s better tailored to the patient—and help put an end to the practice of doctors defaulting to the drugs with which they’re most comfortable.
Artificial intelligence may also shape the next generation of cancer drugs.
It’s a field ripe for disruption: drug development is time-consuming, costly, and relies on efficacy tests computer scientists believe can be modeled by algorithms. Rather than relying on trial and error informed by the experience of chemists, as traditional pharmacology has done for most of the pharmaceutical industry’s existence, AI software can analyze drugs that have worked in the past and the physical structure of diseased cells to generate a few dozen compounds that are most likely to work, at a far faster rate and considering more variables than humans in a lab. This allows companies to potentially spend much less time and fewer resources to formulate marketable drugs.
Atomwise, a San Francisco startup founded in 2012, is using deep learning to look at 3D structures of cancer and other diseases at the molecular level to build potential treatments. Its algorithm, AtomNet, is trained on thousands of documents on molecular interactions between biological and synthetic substances. The algorithm generates dozens of compounds it predicts will interact with a given disease; researchers then run them through traditional physical tests to see exactly what the interaction looks like in real life.
This process isn’t likely to generate a cure on the first try—but it can narrow down millions of molecular combinations to those with the best chance of effectiveness. Atomwise has yet to come up with a US Federal Drug Administration-approved treatment, and the technology is still in its early days. Nevertheless, many big players in health care are betting on the promise of the company’s method: Atomise has partnered with a number of medical institutions and pharmacological companies including Merck, the Scripps Research Institute, and the University of Toronto to explore testing the compounds. How the drug was generated doesn’t influence the US FDA’s testing process, however. These compounds must undergo the same rigorous and time-consuming tests any other medicine must pass in the US.
Berg, a Framingham, Massachusetts-based company founded in 2006, is taking a similar computational approach, but more specifically focused on cancer. The company is currently in FDA phase I and II trials for a molecule, BPM 31510, that could help the body fight glioblastomas, aggressive brain tumors. In these cases, the body loses the ability to tell that rapidly multiplying cells are dangerous—it becomes confused because glioblastoma tumors are comprised of multiple types of cancerous cells. Berg’s compound works to restore this ability, enabling the body to more effectively target and remove affected cells. The drug is now being tested on glioblastoma, as well as another similar type of rare brain cancer called gliosarcoma.
Using AI to find better cancer drugs could also widen the scope of research. When you no longer need a large team of chemists to develop a drug, smaller enterprises can focus on cancer types that don’t get as much attention. Stephen T. Wong, at the Houston Methodist Research Institute in Texas, is trying to develop drugs for rare childhood cancers, which he says are typically overlooked by large pharma companies because there’s not much profit to be made in such a small patient population.
Wong works with the METEOR data warehouse, where (as with CancerLinQ), doctors pool patient data. Coupled with a repository of 10,000 existing drugs, Wong and his coworkers at Houston Methodist are able to make machine learning-based predictions about which known drugs could be effective in treating other diseases. Simply put, the algorithms search for unintended side effects, which might include curing other diseases. A classic example of this is thalidomide, which was once a sedative that got pulled from the market because it caused birth defects, but was proven to cure leprosy and treat some cancers.
This sort of work is representative of the most exciting promise of artificial intelligence-driven health care: Not one algorithm by one startup finding a cure to the disease “cancer,” but instead a decentralization of care and research enabled by the the replicable expertise of AI. Drug generation isn’t just for pharmaceutical companies anymore. Diagnosis can be assisted by an app, and so will no longer only be for the wealthy and able. Planning treatment regimens will take hours, not minutes, giving doctors back valuable time in their days.
“Automation is providing a different way to look at medicine,” Wong says. “Basically the whole idea is to make it faster, cheaper, and better.”