CHICAGO – With growing knowledge of the molecular underpinnings of cancer and access to more precision oncology drugs, academic and community cancer centers alike are increasingly open to using artificial intelligence to help them figure out how to treat their patients.
In research presented at the 2019 American Society of Clinical Oncology annual meeting here earlier this month, Manipal Hospitals of Bangalore, India, reported that recommendations from IBM Watson for Oncology prompted the organization's multidisciplinary tumor board to change its treatment decisions nearly 14 percent of the time. The organization studied 1,000 patients with breast, lung, or colorectal cancer over a period of two years.
In cases where Watson changed clinicians' minds, the computer provided evidence for a newer treatment 55 percent of the time. Suggestions of a "more personalized" option based on each patient's unique clinical characteristics, accounted for 30 percent of changes, according to SP Somashekhar, chairman of surgical oncology at Manipal Hospitals. About 15 percent of decisions to go a different direction stemmed from insights from genotypic and phenotypic data and "evolving clinical experiences," he added.
"Even in a tertiary care functioning [multidisciplinary tumor board], with a group of top oncologists, it adds [more than] 13 percent more value of decision making, which helps us to arrive at a personalized, evidence-based recommendation for that patient," said Somashekhar.
Manipal has a busy academic cancer center that gets 60 to 80 patients a day. Somashekhar said that the impact would be higher in community clinics that lack subspecialized oncologists.
Andrew Norden, a former lead physician for oncology and genomics at IBM Watson Health and now chief medical officer of Cota Healthcare, agreed that artificial intelligence and machine learning will be particularly useful for community oncologists, who tend to treat all kinds of cancer rather than focusing on one type like their academic counterparts.
"Cancer is becoming an increasingly complex set of diseases in the sense that there are all these new molecular targets that are helping to define clinically important subtypes of cancer and for which we have a growing number of treatment options," said Norden, an attending neuro-oncologist at Dana-Farber Cancer Institute in Boston. "It's getting harder and harder for an oncologist to keep up."
As datasets grow in size, AI helps identify sophisticated patterns that the human brain and some older analytics technologies might miss. "Advanced analytics and artificial intelligence have the potential to be increasingly helpful in ensuring that doctors are aware of these clinically meaningful mutations and what therapeutic options may be associated with those," Norden said.
He noted that while community oncologists now have better access to genetic tests, there is still a shortage of interpretation and counseling services. AI systems can help community-based practices develop precision treatment plans to reduce unnecessary treatment and toxicity, and ultimately improve patient outcomes.
However, according to Norden, medicine so far has not done a good job measuring patient outcomes. "Frankly, I think there's a ton of value today in simply curating medical records such that you have a queryable dataset and then looking at treatment plans, outcomes, and costs," he said.
Still, machine learning is not plug-and-play, as it does require some human intervention to create suitable datasets.
"It takes a lot of work from a technology and human standpoint to clean up a medical record to the point that there exists an analyzable dataset," Norden said. "Machine learning will be a critical technique, but I think that there are lots of benefits that one can achieve simply from creating the dataset, making outcomes available, and doing more standard analytics."
He believes that natural-language processing will be important for at least another decade because physicians still largely document patient encounters in unstructured text. It simply is easier for physicians to dictate notes than it is to enter data into an electronic health record in a structured format.
While there is a need for techniques that are better at extracting information from unstructured text, Norden said that doctors will only start entering formatted data when it becomes easier for them to do that instead of dictating notes. Or the incentive structure must change so it "becomes a very bad idea for doctors to stick to the old way of entering data," he said.
Doctors should ultimately come to view AI as a tool that can save them time by letting the computer perform "menial tasks" like searching literature, noted Nathan Levitan, the current chief medical officer for oncology and genomics at IBM Watson Health. "It's the best way to use AI," he said.
Yet, no matter how powerful the computers and algorithms are today, artificial intelligence is no panacea for medicine. Watson has certainly had to make some adjustments in the clinical space. In July 2018, Stat News reported that IBM's supercomputer sometimes gives inappropriate cancer treatment advice.
IBM documents obtained by Stat reportedly showed that the system recommended that a patient with lung cancer and severe bleeding be treated with chemotherapy and bevacizumab, even though bevacizumab can lead to hemorrhaging. The documents appeared to attribute the incorrect recommendation to the training Watson received from IBM engineers and Memorial Sloan Kettering Cancer Center physicians on hypothetical cases, according to the report.
MSK said that it believes this case was part of system testing and the recommendation was not given to a patient.
In a Journal of the American Medical Association article published last month, University of Pennsylvania oncologist and medical ethicist Ezekiel Emanuel and hospitalist pioneer Robert Wachter of the University of California, San Francisco, wrote that like any other technology, AI is merely a tool that has to be applied properly.
"While AI has been responsible for some stunning advances, particularly in the area of visual pattern recognition, a major challenge will be in converting AI-derived predictions or recommendations into effective action," such as in patient care, Emanuel and Wachter wrote.
They said that earlier forms of clinical decision support have not had the desired effect for many reasons, including inadequate user experiences and an excessive amount of warnings that delivered "alert fatigue," causing clinicians to tune out or disable the technology. Emanuel and Wachter explained that clinicians will not alter their routines to adhere more closely to standards of care simply from better computerized predictions.
Still, AI can have a role, as long as the focus is on mechanisms to put the technology into practice in support of "positive behavioral changes," they said.
At ASCO, Philips unveiled a collaboration with Dana-Farber on the Oncology Pathways tool, an optional component of the Philips' IntelliSpace platform. Using workflows developed at Dana-Farber, the partners are marketing Oncology Pathways to community clinics.
This is the next step of a year-old partnership to deliver genomic- and diagnostic-centric clinical decision support to clinicians through EHRs, following best practices developed at the Boston cancer center.
Dana-Farber curates clinical oncology pathways that David Jackman, Dana-Farber's medical director of clinical pathways, said try to reflect the complexity and granularity inherent in cancer care. "Countless person-hours have gone into trying to put the right recommendations in the right places," he said.
The pathways advise clinicians at the time of diagnosis or during the treatment process about appropriate biomarker testing, such as for mutations, fusions, PD-L1 expression, or MGMT methylation.
For example, Jackman said, the pathways for newly diagnosed metastatic, non-squamous, non-small-cell lung cancer, recommend that tumors be tested for mutations in EGFR and BRAF, as well as for rearrangements and fusions in ALK, ROS1, and NTRK. "Broad next-generation sequencing is strongly encouraged," Jackman added.
Should biomarker testing turn up actionable results, the Dana-Farber pathways provide appropriate treatment recommendations including dosing, schedule, and literature citations supporting the decision.
The pathways also point out treatments that are contraindicated for various biomarkers.
For instance, in patients with chronic myeloid leukemia who are on first-line imatinib therapy, "the finding of a Y253H mutation not only suggests to the user to consider dasatinib or bosutinib, but it also notes that this mutation is associated with resistance to nilotinib," he said.
The Philips technology with the Dana-Farber pathways is commercially available now. A Philips spokesperson said that the IntellSpace Oncology Pathways system would be going live at Dana-Farber "soon," but did not offer any specific timeline.
"We see this as a learning system on so many levels," Jackman said. Dana-Farber convenes its experts several times a year to discuss new medical literature and research, as well as to examine the data over the preceding few months of how often patients received treatments recommended by the pathways and why physicians decided to diverge.
"Bringing this kind of information to our practice has been incredibly useful. It helps us to oversee our practice," including when and why physicians decided to try something different, Jackman said. "It also tells us how robust our pathways are. If people are consistently on our pathways, we take that as our pathways really looked into the situation pretty well and we covered the granularity of what we expected to see."
There are more formal ways of learning as well. Dana-Farber published in the Journal of Oncology Practice in March 2017 a review of a year's worth of patient outcomes in metastatic non-small-cell lung cancer before and after the pathways went live.
"There was a significant cost savings after the pathways went live," mostly from decreased chemotherapy usage and costs, Jackman explained. The medical team that developed the pathways did not see additional benefit from certain expensive drugs. Although the outcomes "trended toward better," he said, the findings were not statistically significant.
Reducing reliance on chemotherapy and other potentially toxic treatments is another popular goal for AI in oncology these days.
In a presentation to the 2018 Machine Learning for Healthcare conference at Stanford University, Computer scientist Pratik Shah, a principal investigator at the MIT Media Lab, discussed machine-learning techniques for reducing dosing of chemotherapy and radiotherapy for glioblastoma.
A model he presented involved a "self-learning" algorithm that studies treatment regimens and adjusts doses to optimize treatment for minimal dosing to achieve similar tumor size reduction than in a control group based on differences in original tumor size, medical histories, genetic profiles, and biomarkers.
The system also considered pharmacogenomics. "The drug concentration is assumed to induce DNA damage in both proliferative and quiescent tissue through linear functions (damages in proliferative and quiescent tissue were factored in and used for learning)," according to Shah's presentation.
In a simulated trial of 50 patients, the system cut potency of radiation and chemo by 25 percent to 50 percent of all doses without reducing "tumor-shrinking potential," Shah said. In many cases, it recommended skipping doses completely, to as little as twice a year instead of monthly.
This technique, called reinforced learning, borrows from behavioral psychology by giving more weight to activities that can achieve the desired outcome.
Shah and colleagues applied reinforced learning to glioblastoma treatments involving a combination of temozolomide and the drug cocktail of procarbazine, lomustine, and vincristine (PVC) to achieve their results. The system gets rewarded or penalized as it takes actions and for outcomes. If the goal is to reduce toxicity, the computer should not recommend repeated dosing to shrink tumors, he said.
"It's slightly more unsupervised than standard machine learning," Shah said. "In reinforced learning, you set the task, but you don't micromanage how the machine learning-AI agent gets there. You give it some flexibility of trying out a variety of things in the dataset or the search space that it has to optimize the outcome you want."
His team is in conversation with hospitals to pilot the models in a clinical study and is in the process of applying to institutional review boards.