It seems almost every day there’s exciting news coming out about the promise of artificial intelligence to cure cancer. Just a few examples of the pace of discovery and the hope attached to these types…
It seems almost every day there’s exciting news coming out about the promise of artificial intelligence to cure cancer. Just a few examples of the pace of discovery and the hope attached to these types of technological developments include a recent study detailing how a machine-learning technique could speed up identification of ER-positive breast cancers, image-based technologies that are learning to spot genetic mutations in lung cancer just based on how samples of the tumor look and AI-based skin cancer diagnostic tools that outperformed experienced pathologists. With all these advances, hope abounds that AI might be the future of cancer detection, identification and treatment.
Several startups have rushed to capitalize on this forward motion, leading to other recent news reports suggesting artificial intelligence systems designed to help doctors diagnose cancer and predict outcomes may not be ready for prime time. Recent reports have noted that prominent AI-based cancer detection system IBM Watson for Oncology was trained on synthetic patient records, rather than real cases as was advertised, and that it generated “unsafe and incorrect treatment recommendations” according to leaked internal documents.
This mixed bag of reports about what’s going on in the AI-for-cancer space — a potentially huge market that could completely disrupt how cancer care is delivered — indicates that for the moment, the promise of AI seems very bright, but there is still a long way to go before such technologies will become commonplace in the clinic and an even vaster gulf to cross before we enter an era where machines might replace human diagnosticians and prognosticators.
A World of Opportunities
The future role of artificial intelligence in health care — and within the diagnosis and treatment of cancer specifically — will likely take a number of forms. From diagnosing a specific form of cancer to determining which treatment approach may best treat that particular case, AI promises to improve personalization of cancer care and help people deal with the disease with a higher quality of life and fewer side effects.
One specific technology that has shown a lot of promise uses a Google AI tool to identify genetic mutations in a tumor based on images of lung cancer tissue collected during a biopsy. Aristotelis Tsirigos, associate professor of pathology and director of the Applied Bioinformatics Laboratories at the New York University School of Medicine, led the study that showed this particular program may outperform pathologists and explains that this technology mimics how the human brain works. Called neuronal networks or deep learning, this approach to AI was first proposed in 1944 and has expanded significantly over the past decade or so. In a nutshell, these types of AI systems rely on an algorithm — a complex computational program — that crunches vast amounts of data fed into the system. Over time, it learns all the imported data and in some cases can even begin to infer things it hasn’t directly been taught — hence the term deep or machine learning.
Tsirigos says the term “deep” refers to the many layers of data processing involved in these complex computational analyses. “You can think of it conceptually as every layer building on the previous one.” The first layer sees that there’s an image and “identifies some very basic features such as the colors and the lines maybe.” The second layer identifies additional features of that first set of features and subsequent layers identify additional details to create a “structure of potential features you can extract from the image.”
Then, these various features are given weights — an assignment of how important they are to solving the question at hand. For example, does a certain color or shape have meaning that needs to be given more value than another color or shape? This builds the neuronal network, which “simulates human thinking because we’re thinking of features that are important for decisions that we make. How much weight are we putting on every feature that we, as humans, think is important to make a final decision? Some things weigh more, some things weigh less, so this is the process when we train a neural network,” he says.
This training period, during which images are fed into the algorithm, requires the efforts of a human teacher, but over time, as more data sets are imported, the algorithm “learns” what to look for and how to make its own decisions and assumptions in such a way that it actually starts to figure things out that the human brain simply doesn’t have the capacity to process.
“Over the last 10 years, a lot of technology, most notably genetic sequencing, came out, meaning a lot of data sets were generated. So now it’s as if we’re harvesting [that data] using AI,” and analyzing it in deeper, cross-referential ways that can reveal more detail and information than the first, human-led pass through the data was able to produce. As more raw data sets became available, “AI became relevant because of the size of the data. With small sizes of data, you don’t need AI — you can just use regular statistics. But if you have big data, then you have to use AI to reap the benefits.”
Tsirigos says that eventually, our current data production methods will lead to a plateau in what AI is able to do unless we can “generate more data or more data modalities. I think what’s going to happen, maybe in a few years we’re going to say AI explored all the data sets we have now,” and we need new and better ways of deriving more meaningful data. He says this field will have to develop “hand in hand” with evolutions in laboratory procedures to gather that data.
If and when data production technologies plateau, Tsirigos says breakthroughs in combined data approaches may still yield results or offer better answers to different questions, such as what’s the best targeted treatment for a specific case of cancer. Exploring all the potential drug combination permutations is another possible future area of study that could provide better insight to best treatment options and patient outcomes.
“It’s one thing to be able to make a diagnosis or even predict mutations, but the real question is: Can you predict clinical outcome, meaning, for example, responses to specific therapies and precision medicine or survival?” These are important questions that Tsirigos and his team are working on right now, and the path forward in answering some of these questions may lie in combining approaches and evolutions in how data sets are generated. “We’re not going to be limited to one data type. The ideal scenario is you have a lot of data of different types for a given patient and you combine all this data into your model to predict” how they are going to fare on a certain drug therapy. This data will likely include imaging data, genetic sequencing data that offers information about genetic mutations, which can then be correlated to prediction of prognosis based on those genetic mutations. “There’s no reason why we shouldn’t combine all the information we have from sequencing and imaging, MRI and blood testing. Everything we can integrate can be useful” in arriving at a more precise care plan for individual patients, he says.
Improving Quality of Life
Artificial intelligence may also have applications in helping patients deal with side effects of treatment for cancer. A recent partnership between a major cancer research center on the West Coast and Microsoft hopes to develop an AI-driven device that will help keep cancer patients who are receiving chemotherapy out of the emergency room by detecting subtle signs of a complication before they become a full-blown problem.
Dr. Scott Ramsey, director of the Hutchinson Institute of Cancer Outcomes Research in Seattle, says the partnership aims to leverage the strengths of both the HICOR and Microsoft teams to develop a wearable tracking device or smartphone app that will monitor chemotherapy patients’ side effects and “better manage them before they become bad enough to warrant an emergency room visit. The goal is to combine Fred Hutch’s clinical and data science expertise with Microsoft’s AI/machine learning technology to improve patient health while lowering health care costs in the process.”
It’s a big goal, and the pilot project has only just begun. Ramsey says the team is figuring out which metrics need to be tracked, but vitals such as heart rate, body temperature, energy expenditures and other metabolic information will likely be collected and analyzed for subtle changes that could signal a coming problem.
“Almost all of those problems that people end up in the ER for are toxicities related to chemotherapy that probably could be prevented if the doctors knew that patients were getting in trouble and [could] have the patients take more nausea medicine or hydrate themselves before getting into trouble,” Ramsey says, adding that “among patients who get chemo, about half of them end up in the emergency department or hospital.” The Centers for Disease Control and Prevention estimates that some 650,000 people receive chemotherapy in an outpatient oncology clinic in the U.S. each year. Keeping more of them out of the hospital could save a lot of money and make for a better quality of life for those patients.
Developing a remote monitoring system that will alert doctors to changes in patient health is an exciting idea that could translate to real improvements in care delivery for hundreds of thousands of people — not to mention large cost savings at a national level — if the technology can perform the way the developers hope it will.
Although sometimes it might feel like the future is already here, Ramsey cautions that much of the buzz surrounding artificial intelligence, machine learning and how these technologies may already be revolutionizing the diagnosis and treatment of cancer has yet to pan out as much more than hype. There’s still a lot of work to be done before we can reasonably expect machines to make diagnoses or decisions about the best course of care.
“We have no illusions that it’s going to be easy or that we know all the answers,” Ramsey says. “But that’s part of the fun — figuring it all out.” Putting the pieces together and allowing the technology to catch up with the concept will take time. “It’s not that there’s going to be some sweeping revolution. We’re going to have wins here and there, and I believe getting chemotherapy patients out of the ER is achievable. I’m as uncertain as anyone about all the things we need to understand,” but with steady, methodical work, real patient gains may become achievable. “I think these technologies will help doctors and patients in the long run.”