Imagine you’re an oncologist practicing in a community hospital in, say, North Dakota. Because our evolving understanding is that cancer is essentially a genetic disease, you’ve sent a biopsy of a patient’s lung tumor to a lab for genomic analysis and have just received the results. You have to figure out what treatment is best for the patient.
What does that mean in this age of information and genomics? Well, there are at least 250 academic journals related to oncology. These journals publish tens of thousands of papers each year. New, targeted drugs are being approved for cancer all the time — at least half a dozen new therapies since January 2015. Dozens, even hundreds, more are being tested in clinical trials all over the world. Every month, geneticists identify new tumor mutations that may be drivers of cancer. Each patient’s tumor may have dozens of mutations that might affect how the patient responds to treatment.
If that sounds overwhelming, that’s because it is. Even at big, prestigious academic medical centers, where perhaps 10 percent of oncologists practice, doctors find it challenging. How can even a very smart doctor with decades of experience keep abreast of it all?
“Medicine is evolving so fast, especially in cancer, where genomics is revolutionizing treatment paradigms,” says Lynda Chin, the chair of Genomic Medicine and the scientific director of the Institute for Applied Cancer Science at the University of Texas MD Anderson Cancer Center in Houston. “So how do we get the new knowledge to all of the practicing oncologists, because there are simply not enough hours in a day to read all the publications and literature?”
Digital Medical Assistants
For more and more doctors, it is becoming possible to lean on a digital assistant, a computing system that links to massive amounts of data in the cloud and identifies studies, clinical trials, and treatments specific to the particular mutations present in a patient. The system “curates” all this data and presents the doctor with a report outlining various options. The physician, in consultation with the patient, then makes the final call on a treatment plan.
Many companies are working to create digital medical assistants that help doctors deal with the avalanche of information (see sidebar “Watson’s Competition”), but perhaps the best known are the health applications being developed by IBM’s Watson system, the computer that trounced two champions on the Jeopardy! quiz show in 2011.
It’s still too early to showcase patients who have been “cured” or had more effective treatments because of these tools, but experts say those stories should come in the next year or two.
These systems grow out of an academic discipline called artificial intelligence, which attempts to make computers approximate human intelligence. Marketers like to tout the new systems as “cognitive computing”— computers that really think like humans. But many experts say we’re a long way from that.
How Watson learns
On Jeopardy! Watson had to understand an answer and then infer the question linked to that answer, after scanning through massive amounts of data. To do this successfully, Watson had to retrieve information and synthesize it, use “natural language” (the way people actually talk), reason, learn from mistakes, and communicate the results in a form that humans can understand.
That’s a big deal. Questions and answers have been a central issue for programmers since the dawn of computers during World War II. But it’s not the same as a doctor using knowledge, the experience of decades, and intuition to treat a patient, experts say.
“What Watson lacks completely is clinical judgment. Watson cannot go to clinical judgment school,” explains Charis Eng, the chairman and founding director of the Lerner Research Institute’s Genomic Medicine at the Cleveland Clinic. “That’s the gut feeling of a doctor standing in front of a patient — possibly derived from long experience — that something’s just not quite right.”
All computers solve problems according to various rules or instructions, called math algorithms, which are what most of us call programs. You can think of these as very sophisticated versions of the logic sequence “If A and B are true, then C is true.”
For the Jeopardy! challenge, Watson bundled more than 100 of these logic techniques to identify sources, learned how to rank the reliability of those sources, and then came up with correct answers.
Once Watson’s team came up with this logic architecture, it started the “machine learning” phase: teaching the program to succeed in a game show context and analyzing previous Jeopardy! matches and the performance of various champions. To do this, they created specialized rules relevant to success on Jeopardy! Then, the IBM team loaded the computer with the information from encyclopedias, dictionaries, thesauruses, news articles, literary works, and more.
After success on Jeopardy! four years ago, IBM began working on ways to make Watson marketable. The idea is to marry the stronger reasoning power of third-generation computing with the massive amounts of information that are now stored in the cloud. This “narrow AI” doesn’t try to replicate a whole human brain; rather, it tries to give the computer very deep knowledge in a narrow field and then create programs to facilitate analysis and reasoning in that field.
IBM has launched Watson initiatives in several industries, including finance, marketing, energy, social media, and food, claiming hundreds of clients across six continents. But the company has been most active in the health field, with projects that may eventually match patients with clinical trials, help pharmacists keep up to date, glean useful insights from the massive amount of data in electronic medical records, and improve surgical and diabetes care.
Watson in Cancer Centers
Watson has perhaps made the most progress in making personalized cancer care more of a reality. It hasn’t been a steady, easy path; earnings are not as stupendous as originally predicted by the company; and there have been growing pains as the computer team learns to work in various professions.
Computer engineers do not necessarily speak the same language as oncologists or endocrinologists. Remember that each time Watson enters a new field, it has to go through deep learning all over again. It has to go to culinary school, business school, or whatever. The computer has to absorb new rules that approximate how a professional in that field thinks and learn the base knowledge of the discipline. It has begun to make progress in cancer care and genomics.
Three top cancer care hospitals are working to build oncology advising tools with Watson: the Cleveland Clinic in Ohio, MD Anderson Cancer Center in Houston, and Memorial Sloan Kettering Cancer Center in New York. It is also partnering with the New York Genome Center, creating a prototype designed specifically to analyze genomic research and help oncologists use that information to make treatment decisions.
“If you go through the usual channels, from the time you send the block [of patient tissue] to the genomic testing lab, it could be two, three, four months” before the lab responds, explains Eng of the Cleveland Clinic, where they’re doing a pilot test with a set of patients to see how Watson can analyze changes in a cancer cell. “For people with metastatic disease, that’s a bit long. When patients have six months to live, they don’t have three months to wait for a test. Watson would do this analysis in three or four hours, or less.”
“The system first takes the data from a patient’s electronic medical record,” explains Steve Harvey, vice president of the IBM Watson Group. “Then, with that particular case in mind, it searches all the academic literature, all the possibly relevant genetic mutations, all the clinical trials that target these particular mutations. The system then presents the oncologist with treatment alternatives, explaining how it came to these conclusions and where the doctor can find the original data.”
“In a nutshell, this system is ideally contrived to get information about patients, about the meaning of genetic changes, and then to put it in context,” explains Mark Kris, the lead physician for IBM Watson Oncology at Memorial Sloan Kettering.
At MD Anderson, a tool called the Oncology Expert Advisor (OEA) uses Watson’s core language analytic capabilities to take in care guidelines and clinical literature and read patients’ medical records. Then, cancer experts at MD Anderson train the OEA to suggest treatment options based on the patient’s profile and cancer knowledge.
“It’s like a lawyer having a paralegal,” explains Chin. “You send the paralegal to the library, [have them] pull out all relevant cases, and [they] give you the cases to make your argument. We design and train the OEA system to do just that, to bring MD Anderson to all practicing general oncologists so they make decisions for their patients based on the most up-to-date cancer knowledge.”
So far, the MD Anderson team has developed and tested two OEA solutions, one for acute leukemia and one focusing on molecularly targeted therapies in controlled clinical settings. They are working on a comprehensive multidisciplinary lung solution.
In early May, IBM announced that 14 U.S. cancer centers were going to subscribe to a Watson oncology tool that will help bring personalized medicine to cancer patients nationally. None of the partners would discuss the exact financial arrangements, but patients will not be billed directly. Instead, these oncology tools will be lumped in with the rest of the indirect costs of running a hospital, such as keeping the power on or preparing patient meals.
IBM says it is in talks with many cancer centers that want to use Watson’s AI tools. Likewise, partners like MD Anderson, Cleveland Clinic, and Memorial Sloan Kettering will eventually market their tools to smaller community hospitals.
“My hope is that we can get through data quickly and find out what’s meaningful, faster,” says Steven Powell of Sanford Health, a Watson partner that runs medical centers in the upper Midwest. “Now it can take weeks with humans doing this. Ultimately, I don’t think Watson will replace the molecular tumor board or the physicians involved, but it will enhance what they’re able to do.”