As computer-based physician support systems for decision-making in cancer management continue to evolve, “we will come to embrace this as something that liberates us to spend more time on the human aspects of cancer care,” Andrew D. Seidman, MD, told participants at the 2017 Lynn Sage Breast Cancer Symposium, sponsored by the Robert H. Lurie Comprehensive Cancer Center of Northwestern University, Chicago.1
Dr. Seidman is Professor of Medicine at Weill Cornell Medical College, a breast medical oncologist at Memorial Sloan Kettering Cancer Center (MSK), New York, and Senior Breast Cancer Advisor to the MSK-IBM Watson for Oncology Collaboration. “For the past 6 years or so,” he said, “I’ve been involved in a collaboration that has opened my eyes to some unmet needs and how we might leverage the power of the computer to help oncologists make optimal treatment decisions.”
‘Work in Progress’
The MSK-IBM Watson for Oncology Collaboration is very much “a work in progress,” Dr. Seidman said. The progress has come from inputting huge amounts of data from clinical guidelines, journal articles, and clinical trial results for cancers of the breast, lung, and colon, and from “training Watson as we train our fellows—experientially through cases,” Dr. Seidman said. “So we can try to leverage what Watson brings to deliver best care, to improve care through research, and to disseminate that information.”
Work remains to be done in several areas, such as expanding coverage to include a broad variety of cancers, matching patients to clinical trials, and “making sure the patient’s voice is heard, particularly in situations where treatment is not given with curative intent,” Dr. Seidman said.
With the “exponential growth” in clinical information and the increasing importance of molecular details, it is an “uphill battle to get a handle on the vast amount of data needed to optimize patient outcomes,” Dr. Seidman said. While expert guidelines are useful, “these guidelines take us just so far when we are confronted with a specific patient in front of us. There is a level of granularity that broad guidelines don’t get us to.” Computers can take on the problems of integrating vast amounts of data with clinical electronic records and coordinating tasks such as test ordering and insurance authorization.
“Oncology is a relative latecomer to using computer-assisted decision support systems,” Dr. Seidman noted. Such tools have been used for diabetes management since 1998, for the management of dementia since 2002, for sepsis since 2010, and for acute coronary syndrome since 2012.
Watson began as a tool to help guide medical oncologists in prescribing systemic therapy. “Based on early user feedback, we realized that wasn’t going to be sufficient,” Dr. Seidman explained. The multidisciplinary management of breast cancer requires that Watson also be trained about surgical, radiation, and other aspects of treatment and “to put these pieces together to create a uniform multimodality treatment plan,” he added.
“There has been a lot of discussion in the press about whether a tool that’s developed at a single institution can be broadly applicable,” Dr. Seidman noted. He emphasized that Watson “is informed by evidence-based medicine, by literature from top journals, randomized phase III trials when possible, and when not, phase II clinical trials. Clearly, there are circumstances where patients would never have been eligible for a clinical trial. And there, I would submit, we have a very unapologetic bias to teach Watson how to treat such patients based on MSK knowledge and experience.”
Refining Knowledge Along the Way
“Watson is this massive, powerful computing system,” Dr. Seidman noted. “It is a cognitive computer that through an iterative process of machine learning can learn from previous knowledge and can refine this knowledge along the way,” he said. “Computer scientists and oncologists actually sit in a room together and run cases, and when it misfires, we look under the hood and figure out what went wrong and then make alterations.”
Some of the cases used for training are actual cases from MSK, and some are composite or made-up cases with varying patient data and lab values. “It is built on case-based reasoning and deep learning, and it is supplemented by medical literature that is curated by MSK staff,” Dr. Seidman explained.
It is an uphill battle to get a handle on the vast amount of data needed to optimize patient outcomes.— Andrew D. Seidman, MD
“Metastatic breast cancer comprises many, many different diseases, and Watson considers all of those variables, such as disease burden, whether lesions are solitary or diffuse, and whether they are mildly symptomatic, moderately symptomatic, or severely symptomatic,” Dr. Seidman said. The program also takes into account the patient’s organ function and ability to tolerate systemic therapy.
“Watson is not extracting all these attributes right now from your electronic medical record. You need to spend about 5 or 7 minutes telling Watson what it needs to know, and Watson will generate a multidisciplinary treatment plan,” Dr. Seidman said. “It will rank treatment options in much the same way as the Homeland Security Advisory System—green, yellow, and red. Green means the treatment is a highly recommended, yellow or orange treatment is for consideration, and red treatment means ‘don’t do this; it is not recommended or not advised.’”
In response to an audience question about how Watson for Oncology is being used at MSK, Dr. Seidman replied that MSK is currently using the system for various pilots to inform its continued role in developing Watson for Oncology.
“It is being used at other institutions outside of MSK,” he added. “Since we are developers and trainers, we are probably not necessarily the ideal target audience to use it. But there are hospitals in Asia, India, and the United States who are early adaptors. It is important for those early adaptors to provide feedback so that there is a feedback loop as Watson continues to go through this early development.” ■
DISCLOSURE: Dr. Seidman reported no conflicts of interest.