Real-World Data, Real-Time Decisions: AI at the Frontlines of Oncology

As artificial intelligence (AI) continues to advance rapidly, its integration into oncology pharmacy has moved well beyond theoretical applications. From accelerating drug discovery and optimizing precision medicine to enhancing real-time clinical decision-making and remote monitoring, AI is reshaping the way cancer therapies are developed, delivered, and evaluated.

In this interview, Pharmacy Times speaks with Ravi Parikh, MD, MPP, FACP, an attending physician and health policy researcher at Emory Winship Cancer Institute in Atlanta, Georgia, about the evolving role of AI across the oncology care continuum. Parikh presented on this topic at the NCODA Spring Forum in Denver, Colorado, discussing how AI is enabling smarter treatment selections, driving innovations in value-based care, and surfacing complex regulatory and ethical questions that must be addressed to ensure safe, equitable, and evidence-based implementation in clinical practice.

Pharmacy Times: How might Al-driven precision medicine impact the selection of targeted therapies and immunotherapies over the next decade?

Technology integration in health care. Image Credit: © Toowongsa - stock.adobe.com

Ravi Parikh, MD, MPP, FACP: [AI] is being used in drug discovery quite robustly right now, when you think about a standard drug discovery process, which requires extensive screening through drug libraries that can be quite inefficient. What [AI] is being used for is to match existing compounds who essentially do the previously manual work of mining existing drug libraries, whether they be proprietary or public, and using that to match to existing biomarkers that are being discovered quite recently or in a quick, fast pace. And so that automated matching and mining strategy is being used to really accelerate the development of targeted therapies, of drug repurposing for targeted therapies, and even for computational simulation of new compounds that could be the next generation of targeted therapies but just haven't been developed yet.

Pharmacy Times: How can Al enhance clinical decision-making when selecting appropriate cancer therapies for patients?

Parikh: There's a lot that's being out there right now are being developed that are that's trying to translate currently existing guidelines into AI readable formats, such that [AI] algorithms can ingest all of the data that you use to make a treatment decision, for example, and automatically recommend certain decisions at the point of care, say, for example, at the time an oncologist prescribes a drug, or at the time a pharmacist enters in a particular treatment regimen. That's beginning to be really robust and even being able to be curated to individual systems or pathways or payer’s formulary so that you can get to a true precision decision for a patient, rather than having to go through the prior authorization process or go through the uncertainty of trying to select which of many different therapies might be the best one for your patient.

Pharmacy Times: How can Al improve medication adherence and patient monitoring for oral oncolytics?

Parikh: The emergence of AI has really been coming alongside the emergence of something called remote patient monitoring, and so the merging of the 2 has really opened up this new possibility for improving medication adherence and monitoring for fills and overdue refills of oral oncolytics. So just, as an example, these so called smart pill boxes can automatically measure not only whether you've taken a drug or not, but also whether you are taking it at the right time during the day, or monitor what times you are taking it at a given timing of the day, and such that if there's a symptom that's reported through an electronic pro capture system, for example, then there can be an automated AI-based algorithm that automatically recommends a dosing adjustment or a dosing hold concurrent with guidelines, rather than relying on a patient or family member to wait until you can get in touch with a given clinician, and that sort of combination of remote patient monitoring, not just for smart pill boxes, but also for remote symptoms or remote physiologic measurements, combined with AI-based algorithms that work on the back end can really serve to offload these really overburdened triage lines that oncologists and pharmacy staff and others have when trying to respond to patient concerns about these medicines, and over time, I think, result in more precision dosing of these type of therapy strategies. And this isn't now pie in the sky. This is already being able to be built out there with really complicated regimens, like, for example, [capecitabine and temozolomide] for neuroendocrine carcinomas and even for other ones down the line.

Pharmacy Times: What role does Al have in analyzing real-world evidence to guide oncology treatment protocols?

Parikh: So, the emergence of [AI] has also [brought] along the emergence of increasing amounts of real-world evidence, and I would argue that the emergence of AI really wouldn't have been possible had it not been for high quality, real world evidence with granular curation of patient data to be present, because you need good data to build good AI models.

Now, the emergence of relatively real time, only delayed by a few months or so, real world evidence can be combined with AI to assess a few different things. For example, real world comparative effectiveness, analyzing whether patients in the real world respond the same way as they do in clinical trials, and if not, automatically curating different indications so that they match the level of evidence that's coming out there through routine post-market assessment. That's one real possibility of [AI] and being used to generate faster regulatory decisions.

The other way that this could be used is even if it's not restricting or expanding a given indication to help comment on dosing strategies, since AI-based algorithms that are used as part of causal inference platforms can really help to analyze whether a particular dosing strategy or whether a particular dosing change would be beneficial for a patient faster than the years and years it would take to run a clinical trial. So those are just good examples of how we're able to use AI to discover emerging patterns in real world evidence that can result in treatment or regulatory decisions much faster than would normally otherwise occur.

Pharmacy Times: How will Al influence cost-effectiveness analyses and value-based care models in oncology?

Parikh: Well, I can think of a couple of ways that AI will be used to influence cost effectiveness analysis and value-based care models in oncology. And these will get a little bit specific, but one is around the idea of prediction: predicting potentially high cost or high needs patients, since it's often these are the patients that we need to be directing resource constrained but effective interventions like care management programs. So now they're beginning to be AI-type algorithms that are trained off of large claims-based data sets to automatically predict patients even within specific different cancer phenotypes that are likely to be high cost or likely to incur a hospitalization or incur costly end of life utilization, for example, in the coming year to 2 years. And so, merging those algorithms inside value-based care models is really a perfect melding.

There's also beginning to be use of [AI] for coming up with better episode groupings in value-based care models or better risk adjustment strategies from groups like commercial payers or from Medicare. And so, we're participating in some of this research as well. So, the use of [AI] in that can really be useful in helping to risk adjust for clinical context appropriately, so that organizations that are making good faith attempts at restricting costs can be rewarded for such rather than being penalized because they don't conform to an outdated risk adjustment algorithm. So those are a couple of ways that AI is being used in that type of policy context.

Pharmacy Times: What are some of the ongoing regulatory and ethical challenges of using Al in oncology pharmacy practice?

Parikh: I can think of 2 to 3 discrete regulatory and ethical challenges of using AI in oncology pharmacy practice. So, the first is that they're beginning to be radiologic/pathology-based, electronic health record–based, genomic-based biomarkers that are derived from AI algorithms to actually help predict whether treatment will be effective or not, and some of these are even making their way into comprehensive cancer guidelines that are used to determine whether we should be giving a certain treatment or not, or a certain intensity of a treatment or not. These are potentially challenging, though, because the way that the FDA is sort of approaching these AI-based biomarkers is a little bit different than the way we would traditionally be approaching them. The FDA largely regulates AI as a software as a medical device, and so some of the requirements for phase 1 through 3 testing or large scale testing, they're a little bit waived in those contexts, and that can result in some potentially questionable AI algorithms being allowed out there, sometimes without patients or doctors even knowing about them. So, I think we need a much more transparent, uniform, standardized process for these so called AI biomarkers to be getting out there.

The second is around ethical challenges. So, when we talk about using AI algorithms, we have to realize those are all built on patient data, and oftentimes these are data sources that are not part of a clinical trial or some sort of prospective cohort study where patients would consent to their information being used. They're being built on real world data that patients have no idea that they're being used or not. And so, these challenges around data privacy, along with certain other challenges around bias and data sets and training data sets and reliability and trustworthiness of the underlying algorithms are real ethical challenges that we have to take on before we start using these to actually make therapeutic decisions in the pharmacy community.

Pharmacy Times: What are the limitations and risks of relying on Al for oncology treatment recommendations?

Parikh: One of the biggest limitations to relying on AI for oncology treatment recommendations is the fact that sometimes these AI tools have not been trained on the right data sets. So, for example, if I was going to rely on an AI making an oncology treatment recommendation for, say, adding chemotherapy in stage IV non–small cell lung cancer, I would want a relatively unbiased data set, potentially from a clinical trial that was comparing use of a chemotherapy containing regimen against a non-chemotherapy containing regimen, so that I could come up with an unbiased model, because that patient allocation to treatment was being based on a randomized assessment. Usually, these AI tools are not being trained on such clean data. They're being trained on somewhat dirty, real world data that has variable missingness, variable follow up times, and a lot of confounding in which patients are getting which potential treatments, and the AI largely bakes in all of those biases into its model. And so sometimes it can result in a really misinformed prediction that, if deployed in a decision support setting, could really influence incorrectly in my treatment decision. So, I think we need to have some increasing standards for the data sets that these algorithms are built on.

The second risk of relying on AI is liability related. Ultimately, in our current liability framework, it's clinicians or pharmacists who hold liability for making an incorrect decision, not the AI. The AI developers are very rarely being sued, even in very significant cases of patient injury, because it's usually any decision is being routed through the end user clinician. So, in that setting, we really need to be careful on so called automation bias, blindly relying on the AI to help make our treatment decision and instead regard AI as one of many data points that we should be melding into our clinical intuition to decide whether a given treatment is worthwhile to give or not.