The role of artificial intelligence (AI) in cancer care was discussed in multiple sessions across this year’s ASCO Annual Meeting. We focus here on two applications.
Marina De Brot, MD, PhD, presented the use of AI-assistance software for HER2-low and HER2-ultralow immunohistochemistry (IHC) interpretation training to improve diagnostic accuracy of pathologists and expand patients’ eligibility for HER2-targeted treatment (Abstract 1014).
Accurate HER2 IHC scoring is essential for guiding breast cancer treatment. HER2-ultralow and HER2-low subcategories of breast cancer are challenging to evaluate and may be misinterpreted as HER-null. In collaboration with MindPeak, AstraZeneca developed an online training platform integrated with AI to support HER2 virtual training masterclasses.
This study determined the concordance and enhancement of scoring accuracy in HER2 IHC assessments with versus without AI-assistance through targeted training of pathologists and AI utilization.
Across 1,940 readings accuracy and concordance compared to the central reference score increased with versus without AI: Pathologists achieved an accuracy of 89.1% with reference scores without AI, compared to 96.1% with AI-assistance. Enhancement of accuracy and concordance for HER2 clinical categories (null, ultralow, low, positive): accuracy improved from 90.1% without AI to 95.0% with AI.
AI support raises sensitivity across HER2 null/ultralow/low expression classifications in evaluating HER2-null from 54% to 88.2%, for HER2-ultralow from 50% to 93.2, and for HER2-low breast cancer, sensitivity from 78.6% to 90.4% without and with AI support, respectively.
HER2-null misinterpretation decreased from 45% without to 11.7% with AI support; HER2-ultralow underscoring decreased from 30.5% without to 4.5% with AI assistance. AI reduced misclassification of HER2-low and HER2-ultralow cases as HER2-null by 24.4%, potentially enabling more patients to have access to HER2-directing ADC therapy.
Next steps include more master classes for additional countries and pathologists, building a global, unified international database to map scoring gaps and guide AI-driven training solutions, and a multicenter implementation study embedding the AI tool in routine diagnostics to measure downstream clinical effects, including changes in treatment allocation and time-to-therapy for patients with HER2-low and HER2-ultralow breast cancers.
The much-hyped, newly unveiled ASCO Guidelines Assistant (GA) is an interactive AI tool developed in a collaboration of ASCO and Google Cloud. Available only to ASCO members, it is meant to improve navigation of ASCO’s clinical practice guidelines, which users have found difficult to locate and time-consuming to use.
Thomas Kurian, Google Cloud CEO, emphasized that the model used to develop ASCO GA relies only guideline content, provides citations indicating where in the guidelines search answers are located, and does not search for information on the internet.
Clifford A. Hudis, MD, ASCO CEO, guarantees that users of the ASCO GA will find mistakes. “It is a guideline discovery tool, not a clinical decision tool. It is meant to get people to the guideline to follow,” he said. “And you shouldn't trust, especially at the beginning, anything on the face of it.” ASCO is aware that the GA may provide erroneous responses or point to the wrong guideline and requests users to identify errors so their dedicated team can fix them.
Dr. Lederman has no conflicts of interest to report.
Illustration by Diana Connolly
