Your AI-Trained Oncology Knowledge Connection!
Sandip P. Patel, MD, discusses how AI can be used to improve clinical trial design and accrual.
As the use of artificial intelligence (AI) in oncology accelerates at a rapid pace, an area of particular interest for investigators is how this technology can be used to optimize clinical trial design and execution, according to Sandip P. Patel, MD.
“AI is here in cancer care, and one of the areas of interest [for AI] is in clinical trials,” Patel, a medical oncologist and professor in the Department of Medicine at the University of California San Diego Health, said in an interview with OncLive®. “Clinical trials [involve] multiple processes related to the screening of patients, radiology measurements for tumor sizes, [and collecting] clinical data. [We are] thinking about AI tools that can help us in those venues.”
During the
Findings from a review published in Current Opinion in Urology summarized the current ability of large language models (LLMs) to assist with clinical trial matching.2 Data from the review showed that TrialGPT, which uses ChatGPT-3.5, ChatGPT-4.0, and MedCPT, was able to retrieve 90% of relevant trials from a 6% initial collection. The matching accuracy was rated at 0.873. Furthermore, ChatGPT-4.0 alone displayed an 88% agreement with initial physician assessment.
The study authors noted that automated clinical trial matching could improve accessibility for patients, as they would be less reliant on providers to match them with ongoing studies. Moreover, oncologists would no longer be required to keep detailed track of eligibility criteria with this approach, potentially reducing their burden and mental load.
“There are many friction points in clinical trials, starting at accrual,” Patel explained. “[We must] find patients who meet multiple eligibility criteria, sometimes including specific genomic or molecular aberrations. It’s important for us to have tools that can help screen patients for these clinical trials. Clinical trial matching is one of the AI tools that we see utilized.”
In his presentation, Patel noted that AI-based LLMs in oncology are far from perfect. Findings from an analysis published in JAMA Oncology revealed that an LLM tasked with offering treatment recommendations for patients with stage I breast cancer offered at least 1 potential treatment that was concordant with National Comprehensive Cancer Network (NCCN) guidelines.3 However, 34.3% of these outputs (n = 35 of 102) also recommended 1 or more treatments that were nonconcordant with NCCN guidelines. The LLM also produced hallucinated responses, defined as responses that were not part of any recommended treatment, in 12.5% of outputs (n = 13 of 104).
“A lot of clinical data are collected [throughout the duration of] a study, [such as] lab values, adverse effects, and imaging assessments,” Patel said. “Having an AI assist with transcribing or manually editing electronic data capture systems, as opposed to a human, is another area [of opportunity]. Increasingly we're going to see [attempts to use] AI to address these friction points head-on, because they're a barrier to us understanding how we can best help our patients in clinic, even though these are back-end systems.”
Racial and ethnic diversity in clinical trials remains an ongoing problem in oncology.4 Machine learning models can be used to address recruitment biases by optimizing cohort selection and inclusion criteria. Moreover, these algorithms can analyze pathological data to screen large multiracial patient cohorts, thereby facilitating improved patient stratification in biomarker-based cancer trials.
“This is an area where we have to be very careful with how we utilize AI [and] not create a digital divide amongst the geographic and demographic divides that often exist in receiving optimal cancer care,” Patel said. “There's a risk that we can have an [AI] model trained on a certain population that that may not be valid [as a] clinical predictor for a different population because we don't know how that population may differ biologically.”
Patel again cautioned that AI tools cannot be used to increase clinical trial representativeness without careful human monitoring. The data used to train AI-based algorithms are usually sourced from trials, electronic health records, and administrative claims. Thus, human supervision is required to ensure that biases are not introduced into the system, and a multi-stakeholder, collaborative approach is needed to ensure that known sources of bias in machine learning—such as evidence/research, expertise/provider, and exclusion/embedded data bias—are excluded.
“As long as we're cognizant about how we utilize these tools, AI can help to [alleviate] some of the problems we have with the digital divide, but we have to be careful and conscious about deploying AI in a way that aims to reduce disparities, not increase them,” Patel said.
