The results from a research partnership between GE HealthCare and Vanderbilt University Medical Center (VUMC) utilizing artificial intelligence (AI) to enable safer and more precise cancer immunotherapies show that the models they developed predict patient responses with 70% to 80% accuracy.
The study was published March 1 in the Journal of Clinical Oncology Clinical Cancer Informatics. To the authors’ knowledge, this approach is the first attempt to design AI models capable of assessing the risks and benefits of immunotherapy using only routinely collected electronic health record data.
A primary advantage of the models used in the study is that inputs are readily available in patients’ medical records, such as diagnosis codes and medication. Only two features — smoking status and number of prior immune checkpoint inhibitor drugs — were drawn from manually collected data. These additional features are easily obtainable by clinicians and could be readily entered into the model.
“We focused primarily on this routinely collected structured data to build predictive models with the goal that these models would be able to be implemented in any clinical setting,” said Travis Osterman, DO, MS, Associate Vice President for Research Informatics, associate chief medical information officer, and director of Cancer Clinical Informatics at Vanderbilt-Ingram Cancer Center.
The researchers retrospectively analyzed and correlated the immunotherapy treatment responses of thousands of deidentified VUMC cancer patients according to demographic, genomic, tumor, cellular, proteomic and imaging data. They designed AI models to predict efficacy outcomes and the likelihood of an individual patient developing an adverse reaction — providing information that may help clinicians select the most appropriate treatment pathway sooner while potentially sparing unnecessary side effects and costs.
Immunotherapies use the immune system to recognize and attack cancer cells and can be more effective than traditional treatments, but response rates are often low, and side effects can be severe.
With the broad availability of input features, the models have the potential for wide deployment and adoption. GE HealthCare is evaluating plans to commercialize, upon securing applicable regulatory authorization, such models for use both in pharmaceutical drug development and for clinical support.
“We aim to partner with pharmaceutical companies, researchers and clinicians to optimize and ultimately apply the AI models in therapy development and in clinical practice,” said Jan Wolber, Global Product Leader – Digital at GE HealthCare’s Pharmaceutical Diagnostics segment. “We want to use AI to personalize predictions and provide decision support for the clinician in determining appropriate therapies.”
The methodology of these AI models is scalable with the long-term potential for use in other care areas, such as neurology or cardiology, the researchers said.
Other Vanderbilt authors on the study are Kathleen Mittendorf, PhD, Michele LeNoue-Newton, PhD, Protiva Rahman, PhD, Cheng Ye, PhD, Neha Jain, PhD, Marilyn Holt, PhD, Douglas Johnson, MD, MSCI, Ben Ho Park, MD, PhD, Christine Micheel, PhD, and Daniel Fabbri, PhD.