Once a topic of sci-fi and fantasy novels, artificial intelligence (AI) is now breaking ground in many fields of science, including oncology. But to successfully put applications of AI technologies to use in clinical practice, careful consideration of when and how these technologies are used must be considered. That's the bottom line Alexander Pearson, MD, PhD, Assistant Professor of Hematology and Oncology at University of Chicago Medicine, makes with his coauthor in a new commentary article published online ahead of print in the journal Cancer (2020; doi.org/10.1002/cncr.33284).
"In this rapidly developing field, there are few established standards, and oncology researchers and providers must educate themselves about emerging AI technology to avoid common pitfalls and ensure responsible use," the authors noted in the article. In an interview with Oncology Times, Pearson elaborated on the prospects of AI in oncology and why this topic is so important right now.
1 What prompted you and your coauthor to write this article now?
"As an oncologist and quantitative scientist, I want to see developments in AI do the maximum amount of good for my patients. In my opinion, a major component of that is by having applied AI research in cancer be led by academic cancer researchers-meaning the teams caring for patients.
"I think academics will add rich, clinically relevant research foci to the terrific work being done by our industry partners.
"In the same way we have seen gene expression profiles integrated into risk stratification and patient care over the last 2 decades, I expect AI-based digital biomarkers to make their way into standard practice in the coming decades."
2 What are some of the biggest opportunities and challenges for AI in oncology?
"AI is changing the landscape of oncology by broadly increasing what type of data can be used for automated, quantitative analysis. A mix of convolutional neural networks and digital processing allows for the rapid extraction of visual data. Furthermore, the flexible architecture of a neural network model means that, in any clinical domain (think risk score or biomarker) where in the past you might include clinical variables or lab values, you might now also include pathology data, radiology imaging data, or a stream of time-stamped lab data. So, models might be applied to all of the above, from prevention and diagnosis to prognosis and treatment selection.
"For example, with AI, we have shown that we can accurately predict the gene expression states of thyroid neoplasms and the microsatellite instability states in gastrointestinal cancers with high accuracy directly from images of the standard pathology sample. In the future, an AI platform using a combination of pathology, imaging, and clinical variables might help a clinician instantly select from a panel of different possible cancer treatments.
"Some of the main barriers to implementation are the availability of high-quality datasets and lack of shared communication between oncologists and computer scientists or machine learning developers.
"As cancer researchers learn how to describe the importance of their research questions and the opportunities for algorithm development to the AI community, along with the right data to answer the questions, I anticipate AI research continuing to accelerate.
"One further shortcoming is the potential for bias or 'overfitting,' meaning that AI models can learn idiosyncrasies of the training dataset as important generalizable features. In order to integrate AI models in clinical workflows and make medical decisions based on their outcomes, rigorous prospective AI biomarker validation is essential (which is no different than with any biomarker)."
3 What's the next step to better utilizing AI in oncology research? What's the next step that will help propel this type of research?
"While the jargon and technical details of AI don't make it easy to break into, the fundamental principles are within reach to most clinicians, so it is worth reading to understand this important emerging technology. Fortunately, the machine learning field in academics has a very open philosophy, and there are many good reviews and resources online.
"Projects that would lend themselves to evaluation with machine learning methods might share some of the following aspects: a large volume of well-organized data, digital (or digitizable) data from multiple streams (i.e., imaging, lab data, medical charts, etc.), and a high potential for clinical impact."