Further Validation Required If AI Models Are to Revolutionize GI Cancer Care

Jacob Shreve, MD, MS

Although artificial intelligence (AI) models have the potential to revolutionize patient care in oncology, rigorous validation of AI models is needed to ensure its successful integration into clinical practice, particularly in the realm of gastrointestinal (GI) cancer, where AI could facilitate more personalized approaches to diagnostics and therapeutic selection, according to Jacob Shreve, MD, MS.

In his presentation at City of Hope’s Annual Advances and Innovations in Endoscopic Oncology and Multidisciplinary Gastrointestinal Cancer Care meeting, Shreve discussed current research and advancements in AI models for cancer care, and the potential impact of leveraging this technology in GI cancers.

“After being promised for decades that AI is going to change medicine, recent trends in computer science, hardware, data, and society are lining up to make it happen,” said Shreve, who is a hematology/oncology fellow at Mayo Clinic, in Rochester, Minnesota. “The floodgates haven’t opened yet, but with a little more rigor and external validation, we’ll start to see AI-models be a regular part of clinical life.”

During an interview with OncLive^®, Shreve highlighted the intersection between AI and clinical medicine, emphasizing the importance of rigor and reproducibility in AI studies to facilitate integration into clinical practice, the potential of multimodal AI models combining various data sources for diagnosis and treatment, and the potential use of AI to address unmet needs in GI cancer through a more personalized medicine approach.

OncLive: Could you discuss some of the key points from your presentation on AI in clinical medicine, specifically detailing your work on a disease agnostic software pipeline combining AI data into cohesive models?

Shreve: My background before medicine was in computer science and AI. Because of that, I feel very motivated to bring my previous career experience into medicine. The majority of my presentation is really going to be introducing AI and how it can affect medicine and how we can improve what we do as clinicians. In terms of our patient care, I want to be able to help educate and promote AI and what it may be able to do in a realistic way, not in an overblown way that you might see in the media.

There are lots of promises regarding AI, but a lot of them are real and about to come to fruition. If you look at the rest of society and the rest of the industry, everybody’s using AI. It’s very helpful and it can really alleviate lots of pain points and different kinds of business practices, the same as with medicine. The problem thus far is that there hasn’t been [enough] rigor and reproducibility with most AI studies to allow it to be integrated into clinical practice, but we’re just around that corner.

The future of AI in medicine is multimodal, taking different types of data and combining it into a cohesive model. I’ll be presenting some of my personal research in which we combined things like automated computer vision analysis of different radiology scans, as well as electronic health record [EHR] data into models that can be used for diagnosis, prognosis, or treatment or efficacy estimation.

What research is currently being conducted to review and validate the utility of AI models in GI cancer, and what needs to be done for these models to be successfully incorporated into clinical practice?

One of the studies that we’ll be presenting is a review article that looks at dozens and dozens of different AI models currently available for GI practice. [This includes] everything from liver fibrosis diagnosis to liver cancer detection, to GI health, and [covers] many topics within the GI [realm]. Some studies are better than others that are reviewed in this article, but all of them are aiming to modify clinical practice using big data in a way that hasn’t been done before. Most of these studies are on the cusp of utility, but they really lack a few things like better external validation, better internal rigor in terms of quality assessment, to be useful to the clinician.

Doctors are very careful people, and they’re slow to change and adopt new technology. [This is] especially [true for] something like AI, the workings of which is a bit opaque, and which requires the person who’s promoting the work and who’s doing the research to go above and beyond to demonstrate that rigor and reliability. I’m interested to see these different practices be used in clinics. It’s become a larger national talking point within medicine.

How can AI models aid precision medicine approaches in addressing unmet needs in GI cancer?

Precision medicine is almost entirely [aimed at addressing] unmet needs when you think about the different demographics of our patient populations and how [despite this], we treat them all the same. People have different attributes in terms of their past medical history, their current health, and their different diseases, but most treatments are still one size fits all. [This is] especially [evident] with cancer care, where there’s a limited number of trials and therapeutics that have been tested. [That] one-size-fits-all approach is not only [seen] with drug selection, but also with drug dosing, follow-up scans, scan interval length, diagnostics and such. With precision medicine we can understand more about each patient and have more accurate predictions [of outcomes]. Maybe a patient shouldn’t get so much of a certain drug, or maybe their scanning interval should be a different length. All those things are going to be improved with precision medicine.

What I’m most interested in seeing in the next year or 2 is that if you use big data in a responsible and sophisticated way within medicine, it unlocks the doors to that kind of precise, patient-oriented care that we’re just on the cusp of. I’m very excited to see that happen. Again, most of that is due to the lack of reliability and external validation [in current studies]. Many of these [approaches] seem to be able to create very good, precise prognoses and diagnoses, but without rigorous external validation, the general clinician population is just not going to adopt it. I would like to see that happen.

Could you expand on some of the other research and/or topics of interest you are excited to see discussed at the City of Hope meeting?

My personal research and my career interests are mostly in precision medicine using AI. There’s an entire panel discussion on precision medicine at the conference that I’m very much interested to see, and the different panelists have different backgrounds and approaches to mostly early detection and precision medicine with cancer care. In terms of GI-based cancer types, it’s going to be very interesting. I pursue precision medicine through a multimodal data method using AI. That’s one way to combine big data and try to find answers to how you can personalize care in a more meaningful way for the betterment of everybody. [It’s for the betterment of] the health care system with less wasted resources, the betterment of the patient with more accurate diagnostics and therapeutics, and better patient outcomes overall. It’s interesting to see everybody else’s perspectives. Precision medicine is such a multidisciplinary endeavor that [requires] a community of specialists to come together and learn from each other’s experiences.

What future research directions in this space are you most interested in?

My very optimistic and forward-thinking self wants to see the culmination of all the different collectible data. There’s an idea that when you’re modeling something with AI, you’re trying to build an artificial or a synthetic version of an actual biological system, [such as a] medical system or a personal health system, which are infinitely complex. The more data that you capture and fold into that model, the closer you get to modeling that real system. If you want to be accurate and sophisticated, you must combine all the different data sources that you can. That’s not only [data from the] EHR, radiology, digital pathology, or genetics; it’s also wearable devices, patient feedback, patient questionnaires, and all these things together. The more sophisticated and the more multimodal a model becomes, there can be pitfalls. There are rigor problems and internal quality problems overfitting all these things, but if you can approach that kind of sophistication, then you’re going to unlock answers that have never been available to us before.

What is your main message for colleagues regarding the importance of AI and this meeting as a whole?

This meeting, especially for myself, is extremely important because of the emphasis on precision medicine and forward-looking nature regarding AI. AI has been promised to clinicians for decades, and it’s just about to happen in a real way because of societal changes and what’s going on with the landscape of computer technology and software development. Everything is in place to allow us to do this. It’s now up to the physicians who are creating these studies to better rigor and better external validation. In just a short amount of time, we’re going to be using AI models to help drive patient care for the betterment of the entire system.

Editor’s note: This interview was conducted prior to City of Hope’s Annual Advances and Innovations in Endoscopic Oncology and Multidisciplinary Gastrointestinal Cancer Care meeting.