The use of big data is revolutionizing many industries, allowing greater insights by drilling into evidence with more detail.
David M. Jackman, MD
The use of big data is revolutionizing many industries, allowing greater insights by drilling into evidence with more detail. Pair that with personalized medicine, where clinicians can tailor a patient’s cancer treatment based on biomarkers, genetic aberrations, or other individual characteristics, and oncologists gain powerful insight that can help predict the best course of treatment for each patient that is based on individual disease characteristics, not just cancer type and stage. Medical centers are also taking unique approaches, but they all have 2 things in common: technology and data.
John F. McDonald, PhD
The Dana-Farber Cancer Institute (DFCI) Clinical Pathways program was conceived in 2012 when the growth of their satellite sites necessitated the development of a system to ensure uniform quality of care throughout the network, said David M. Jackman, MD, medical director of the Pathways program. “We recognize that there are commercial pathways out there,” he said. “But we are a center of excellence. We want to make sure any system we use we are proud to put our name on.” They recruited DFCI physicians to develop Pathways in house, using a commercial vendor for their information technology platform. Program investigators wanted to create an “electronic roadmap” that leads physicians to the best treatments currently available, according to the program’s website.
Physicians using the system are led through important diagnostic criteria, including disease stage, line of therapy, and genomics, before they are directed to the correct branch in the pathway. The system offers pertinent clinical trials, standard treatments for that setting, dose scheduling, and more. The goals are to help physicians avoid checking multiple sources and to give curated recommendations.
A study published in theJournal of Oncology Practicetested whether the Pathways program impacted clinical outcomes and costs.1It included 370 patients with stage IV nonsmall cell lung cancer (NSCLC) receiving treatment at DFCI; the 2 cohorts were defined as those being treated before the rollout of the Pathways program (n = 160; the “prepathways” group) and those receiving care guided by the new technology (n = 210; the “postpathways” group). The prepathways group consisted of patients who got their diagnosis in 2012, which allowed at least 12 months of follow-up before the program was implemented, and the postpathways cohort consisted of patients who received their diagnosis after the NSCLC pathway rollout in January 2014. The Pathways program prioritized clinical options first by efficacy, followed by toxicity and cost.
Researchers found that the clinical outcomes were similar, with median overall survival times of 10.7 months for the prepathways cohort versus 11.2 months for the post-pathways group (P= .08). The study results also showed an unadjusted regression cost savings of $15,993 per patient in the postpathways group (P= .03). The savings were slightly higher, at $17,085, when adjusted for factors such as age, sex, race, and distance from DFCI (P= .01).
The center is currently revamping the Pathways program and moving it to a different healthcare technology vendor, making changes based on what they learned when incorporating it into the physician workflow. This is to ensure that their custom technology gives all of their clinicians the same access to expert recommendations without needing to consult multiple sources and is based on every scenario an oncologist may encounter. They intend to make the system commercially available in 2019 through Philips IntelliSpace.
The new system has 31 medical oncology and 28 radiation oncology pathways. “The pathway portfolio accounts for about 95% of what we or any oncology practice would see, including major and rare cancers,” he said.
Currently, the Pathways program does not incorporate artificial intelligence (AI), although Jackman said that will come in the future. Instead, the DFCI expert panel for each tumor type regularly meets to fine-tune treatment decisions based on the latest publications, data, and FDA approvals and to discuss granular solutions.
Eventually they would like to incorporate AI. “If we are going to take the time to capture the information, we want to learn from ithow physicians behave, how patients do. The process of navigating a pathway provides information to help curate a meaningful database,” Jackman said.
Jackman used the example of a patient with advanced disease and a specific genetic mutation, such as stage IV colon cancer with wild-typeKRASmutation, and said AI can help analyze their response to the medications they are prescribed. Using AI, however, will require caution since the creators of the technology want to ensure the process is transparent.
“There are a lot of concerns about computer learning,” Jackman said. “How and when to incorporate this into Pathways is aspirational.”
Another prominent institution is also using machine learning to predict the best ovarian cancer treatment based on patients’ RNA genomic mutations and past response data from other patients.
Machine learning looks for correlations, said John F. McDonald, PhD, professor in the School of Biological Sciences and director of the Integrated Cancer Research Center at the Georgia Institute of Technology (Georgia Tech) in Atlanta. “The big thing in science is to make predictions; who has cancer, who does not,” he said. That would be a diagnostic use. The other prediction type is for treatmentwhich therapy would best work for a patient.
The first way to do that is to understand the cause-and-effect relationship that underlies the disease. “If you understand what causes cancer, you can design a drug to target the entity causing it,” he said. However, some clinicians do not sufficiently understand the cause-and-effect relationship to design a drug, and that is where machine learning comes in, he said.
“We are looking for tumors with a particular genomic profile, to see how well they have responded to a particular drug,” he said. “That does not presume that we understand anything about the cause-and-effect relationships. It is purely correlative,” he said.
A study published in Scientific Reports, for which McDonald was an investigator, used gene expression profiles from RNA analysis collected from 152 patient records to predict which chemotherapy treatment would produce the best outcome.2The algorithm used 114 test or learning datasets for the model and 38 cases as testing datasets.3The system predicted the correct chemotherapy option, providing the best outcome more than 80% of the time.
To build out the system, metabolites will be analyzed in blood through mass spectroscopy in healthy women to compare with those in women with ovarian cancer. They will then use the computer program to identify correlates, determine disease biomarkers, and eventually develop a diagnostic test.
McDonald said they initially used ovarian cancer tumor data to build the model but then found that they get more accurate models by using RNA data from other cancers that use the same chemotherapy drugs. “We interpret that as cancers are better defined by genomic profiles rather than tissue or origin,” he said. Getting better predictions when using data from all cancers was an unexpected finding.
“The beauty of machine learning is that it allows you an interim solution to the problem of cancer,” McDonald said. It gives insights on what therapies to use for a patient, even if clinicians do not understand why those drugs work. “We can get fairly accurate predictions and we are getting more data every day.” He anticipates that in the near future, every tumor will be profiled on a genomic level, which will provide larger data sets.
McDonald’s lab is using RNA analysis instead of DNA sequencing analysis to genomically profile tumors, as they find it more accurate in identifying gene expression data. They have data for 154 patients, including individual responses to chemotherapy, which comprise less data than they would like at this point. They need both the genomic profile and chemotherapy response for each patient. “That kind of data are very limited,” he said. “Historically, people have not recorded it that way.”
Georgia Tech is making the ovarian cancer treatment prediction algorithm available as an open source decision tool in order to improve it and advance the field. “One of the reasons you cannot get the data you need is that some of the appropriate data are held by drug companies and it is not open source,” he said. “When people develop these algorithms, they are thinking [about] making money on it, so it is proprietary data and nobody knows what is behind the algorithm.”
They put the technology on the software-development platform, GitHub, and invited others to use it and make improvements. “I think it is a much better way to move forward,” he said.
Currently, they are giving the data to clinicians at their institution as an additional diagnostic tool to consider, much like a CT scan. “At this point, clinicians will use the standard-of-care therapies already established,” he said, out of concern for legal implications.
They see the data helping in the short term for patients who fail standard-of-care therapy, where there are no other guidelines for subsequent lines of therapy. The clinician can then use the information to decide on an alternative treatment. “We think, with these algorithms, there is a higher probability of success than randomly trying different drugs.”
When the program has proven itself to a higher degree in the future, McDonald hopes clinicians will use the predictive responses for patients in the frontline setting. “Before that, we need more validation of the accuracy of this approach.”
To create a diagnostic test, they are also now creating metabolic profiles of 1000 patients with ovarian cancer so that machine learning can help predict cancers with a high malignancy potential compared with a low malignancy potential.
Other Computational Approaches
In 2018, Memorial Sloan Kettering Cancer Center formed a Computational Oncology Service focused on bringing together clinical-translational genomics research, including computational biology, cancer biology, and clinical oncology. They plan to use their research and data for quantitative analysis of cancer models, to predict responses and determine better treatments. They have already performed clinical tumor sequencing on more than 20,000 patients to use with machine learning and AI analyses.4
According to McDonald, an emerging area of computer science is network, or graph, theory. It is used in network models that use a hub or a node with many connections. In cancer, researchers can study genes, which serve as the major hubs, and investigate their connections to other genes. That way, they can identify hubs that are attractive targets to break or disrupt the network.
Delivering therapeutics to the cancer cells is also being investigated with computer analyses. Nanotechnologists and chemists from Georgia Tech are working with computers to best target delivery of drugs to tumors, using an integrated approach. “Nothing substantial has come out of it yet, but it looks highly promising for the future,” McDonald said. He believes the big breakthroughs will come from the integration of specialties that have not worked together in the past.
“Healthcare is an industry where we have been reluctant to let ourselves be transformed by data,” said Jackman. “It is very exciting to watch these and many other efforts and try to revolutionize and improve the care we can provide.”