Radiology imaging missing
By Stacy Lawrence
Watson Health is working hard to demonstrate the seriousness of its ongoing efforts in oncology, highlighting its research and customers at the recent American Society of Clinical Oncology (ASCO) conference. The Boston-based IBM business had six presentations at ASCO based on its data and touted that its oncology product is being used at more than 55 cancer centers globally.
But experts in the field of machine learning suggest that the real innovation and value-add for machine learning in oncology is in the analysis of radiology images, which are not addressed by the medical record and literature-based Watson for Oncology. In addition, the lack of access to standardized databases and images, discrepancies in how care is provided and recorded as well as the lack of high quality phenotypic data are all weighing on the useful implementation of machine learning in health care.
The long-term expectation is that machine learning analyses of radiology images will surpass the accuracy of physicians and radiologists, not just provide decision support for them. Ultimately, the Holy Grail for machine learning in boosting individual oncology patient outcomes will be the integrated analysis of all the relevant patient data including genomic testing, pathology slides and radiology images.
MD Anderson stumble
IBM's full-bore ahead approach with Watson for Oncology comes in the wake of a critical article by Forbes early this year in which its relationship with highly regarded cancer center MD Anderson was called into question. The article relied heavily upon a procurement report, an audit conducted by its academic system The University of Texas.
The coverage was critical of the business arrangement, given that the cancer center paid for much of the research-based project itself. But Watson Health notes that its relationship with MD Anderson is ongoing and argues that the report does not reflect on the technology itself or necessarily foreshadow the future.
The press surrounding MD Anderson was very much around a procurement order.
"It was not an assessment of the technology," underscored Louisa Roberts, solution executive at IBM Watson Health, in an interview. "That's very much at the point where MD Anderson can move forward with their Oncology Expert Advisor solution," she told BioWorld MedTech, referring to the joint Watson-MD Anderson product developed under the deal.
OEA achieved almost 90 percent accuracy, despite the fact that the project changed course midway to focus on lung cancer patients instead of those with leukemia.
"We are working with MD Anderson to publish the outcomes of the OEA pilot," noted Watson Health spokesperson Kristi Bond.
IBM's Watson for Oncology that is being trained by Memorial Sloan Kettering, a New York-based MD Anderson rival, is the company's marketed cancer product that addresses treatment for breast, lung, colorectal, cervical, ovarian, gastric and prostate cancer.
The Watson for Oncology data at ASCO were largely focused on demonstrating concordance when recommendations from a hospital's tumor board or an oncologist match-up to the Watson evaluation that is based on the application of the latest literature and MSK training to an individual patient's case. The idea is to ensure that patients, even those physically and/or financially far removed from an elite cancer research center, receive a comparable quality of care.
For example, a cancer center in Bangalore, India, had concordance rates of 96 percent for lung, 81 percent for colon and 93 percent for rectal cancer cases. Another study at a hospital in Bangkok, Thailand, showed an overall concordance rate of 83 percent across multiple cancer types.
Watson aims to integrate its Watson for Genomics, which focuses on genetic tumor analysis, into its oncology offerings. Currently, if genomic information is part of the medical record it is included in the analysis but not necessarily otherwise.
In addition, there is Watson for Clinical Trial Matching, which is being used to better match cancer patients with clinical trials. At ASCO, Highlands Oncology Group and Novartis presented feasibility data showing that it cut the eligibility screening time per patient for a 2,620 patient trial by 78 percent, to 24 minutes from 110 minutes.
Watson Health recently partnered with Novartis in a collaboration to improve breast cancer patient outcomes. The research will use real-world patient data to analyze which combination and treatment sequences lead to the best results.
The full picture
Beyond genomics and clinical trial matching, IBM also has machine learning efforts in pathology and radiology but all of its various projects have yet to be integrated routinely for the benefit of oncology patients. Watson for Oncology may enable health care systems to better manage electronic medical record data, but that's only the tip of the iceberg when it comes to health care and machine learning. The real magic comes with the integration of multiple, and more complex, data sources.
"IBM Watson currently seems to focus most of its efforts in unlocking the knowledge buried in text and the biomedical literature. This is a huge challenge, but much more data is available for patients such as their genome or their medical images, which requires more advanced machine learning models to fuse the data into a single model," Olivier Gevaert, assistant professor of medicine of biomedical informatics research, told BioWorld MedTech.
But EMR-focused efforts, like those of IBM Watson, could help to make electronic medical records more useful and valuable; the implementation of EMRs in recent decades has proven uneven, discouraging physicians and providers.
"There are a lot of efforts to get EMRs implemented, but people aren't satisfied with them. We need to optimize them and machine learning can help," pointed out Ernest Sohn, a chief data scientist with consultancy Booz Allen Hamilton's Data Solutions & Machine Intelligence team.
EMR-based analysis is also limited in so far it often is not a full representation of all the care provided to an individual patient. Sohn also points out that there are a lot of genomic data efforts that are ongoing, but without comparable high-quality phenotypic work that will have little context.
Gevaert notes that deep learning, a more advanced form of machine learning, is well-suited to image analysis which he expects is the most fruitful front for the application of the technology to health care. Gevaert's own research focuses on combining information about the patient as a whole including molecular, cellular and tissue-scale data, a concept known as multiscale data fusion. He sees this as the real key to improving oncology patient outcomes.
His work includes machine learning analyses of CT images that can be used to predict the survival of lung cancer patients that outperformed traditional algorithms and even physicians. He has also done research to apply machine learning to epigenomic data in order to identify new subtypes of head and neck cancer.
Gevaert remains optimistic about the much-hyped potential of machine learning, which he notes is already becoming useful in health care, but sees access to databases to feed it as the real issue right now.
"Machine learning is already making contributions to the U.S. health care system, there are already diagnostics and prognostic tests that are being used to help treat patients that are the product of machine learning research. What we are seeing now is an acceleration of the interest in machine learning and its potential to transform health care. Moreover, the availability of computational resources now vs. a few years ago is also a big factor," said Gevaert.
"Finally, there is the quest for data, the amount of data that is available to study a problem is the most important bottleneck in U.S. health care, as data is not easily shared, transferred and is stored in different formats, that is a real challenge for machine learning," he concluded.
Published June 19, 2017