A UAB researcher is working to establish how artificial intelligence can be used to better understand one of cancer’s most-faced questions. Researchers have spent decades working to address some of cancer care’s most challenging questions, including how to predict what determines whether a patient’s sickness will respond to a treatment. With the help of artificial intelligence frameworks and foundation models, they may be one step closer to finding answers.
Neil Pfister, M.D., Ph.D., an associate scientist at the UAB O’Neal Comprehensive Cancer Center and assistant professor in the Marnix E. Heersink School of Medicine Department of Radiation Oncology who leads the AI in Precision Medicine group, is one researcher working to address this challenge.
With his research and bioinformatics background, Pfister is working to uncover patterns that could improve patient selection, increase drug development and expand precision medicine. Recently, he has worked alongside AI experts to discover how AI models can be used to identify these patterns and how to standardize those models.
Frameworks to identify patterns
Frameworks are created with different architectures, commands and datasets to separate signal from noise and try to identify what is real within a sea of misleading signals. In medicine, AI frameworks can allow researchers to use “electronic medical records and RNA sequencing to derive new insights from very large datasets where there are many more data points than patients,” Pfister said. “We are identifying which patients respond to cancer therapies by distilling real signal from millions of data points per patient, which was not possible before.”
Pfister and collaborators have developed a unique framework shared in their publication “A Deep Learning Framework for Causal Inference in Clinical Trial Design: The CURE AI Large Clinic genomic Foundation Model.”
Clinical trials Uncovering Real Efficacy Artificial Intelligence, CURE AI, is an artificial intelligence framework developed by Numenos that was created to predict how patients might benefit from a new therapy in comparison to standard care for what is being treated. CURE AI uses datasets that include genetic information and medical record information, such as laboratory values, from clinical trial data to formulate predictions.
CURE AI is a “ground-up built architecture” and not an LLM or open source like other well-known AI system models. CURE AI is being used for developing better clinical trials that focus on individual patient biology, so each person receives a treatment that is more likely to be successful for them.
Renal, lung and pediatric cancers are areas of cancer research in which Pfister and his team have worked to gain deeper treatment insights.
“We have the ability to do pan-cancer, a cross-indication selection, which has real-world impact. For example, if we wanted to take a therapy that was just approved in a more common cancer, we now have a method that can take that new and exciting therapy to rare cancers much faster.”
“I am especially interested in being able to utilize information from large, adult cancer studies and bring these therapies to rare cancers faster.”
By utilizing data from existing clinical trials, they can identify what characteristics define treatment response and then apply those same characteristics for patient selection in other cancer types. “We can analyze trials in lung cancer, lock in a predictive test and then predict outcomes in other cancer types, which we have done for immunotherapies,” Pfister said.
Building model architecture to understand insights when there are more data points than patients is something researchers are trying to help by studying different AI frameworks.
“One fundamental aspect that has to be understood between different types of model versions is how to objectively decide which models are better than others,” Pfister said. “And for that, we need really good benchmarking datasets to compare standardized data elements.”
While framework model development can benefit patient treatment, researchers say standardization is vital to the use of AI in medicine. “Some groups are trying to use ‘simulated’ patient data, which is artificial data, to solve problems in patients, because some people believe we do not have enough patient data to begin with,” Pfister said. “However, ‘faking’ data is not going to help improve medical care. What we have shown instead is that, with proper model architecture, we can find real biology even with small patient numbers.”
Benchmarking for success
Benchmarking is vital to artificial intelligence, and it is extremely necessary to how frameworks are being evaluated in cancer and medical research.
AI benchmarking provides a way to assess the efficiency and reliability of AI models by comparing other models to receive the best results. Benchmarks can include diverse types of data and metrics to gauge how successful the models would be in the real-world, clinical setting. Importantly, benchmarking allows different models to be compared to each other.
Pfister is working with major national groups on this effort, such as the National Cancer Institute, the Food and Drug Administration, and MLCommons. Pfister is involved in efforts to understand how to “objectively trust certain models that could lead to clinical implementation or drug development and make better decisions with unlimited data to formulate more refined reasoning.”
Recently, he was a part of NCI’s “Advancing Cancer AI Benchmarks for Real-World Impact” workshop where his work was highlighted by AstraZeneca’s chief of AI for Science Innovation, Jorge Reis-Filho. The workshop’s goal was to begin developing a framework to understand how to benchmark advanced clinical models to evaluate their validity and applicability. Panelists and attendees worked together to brainstorm details regarding actionable next steps for a standardized benchmarking system.
The workshop also highlighted the importance of benchmarking toward oncology applications and practices to help combat the increasing number of cancer patients and to provide the best care possible for those patients, which is crucial for developing new therapeutic strategies for people with cancer.
The work of Pfister and researchers across the field continues to help guide future drug development and more personalized treatment strategies through frameworks and benchmarking to help understand the challenges surrounding cancer. Developing AI benchmarks for medical and cancer care is essential so that different research teams have a standardized way to compare their methods to each other. A secondary advantage for benchmarking is to set goals for teams to strive to surpass, which also serves to motivate researchers and accelerate progress, according to Pfister.