AI Use in Breast Cancer Care Requires Human Oversight: Amrita Basu, PhD

Effective artificial intelligence (AI) for breast cancer care must be embedded within existing workflows and requires human oversight to validate the accuracy of its outputs, Amrita Basu, PhD, of the University of California, San Francisco, told The American Journal of Managed Care^® at the San Antonio Breast Cancer Symposium (SABCS).

She will further explore this topic at SABCS later this week during her presentation, "A Roadmap to Ethical Implementation of AI Tools," which will be delivered during the Thursday afternoon educational session, "Real Impact With Artificial Intelligence."

This transcript has been lightly edited for clarity; captions were auto-generated.

Transcript

What are some of the most impactful AI tools currently being used in breast cancer care?

Some of the imaging tools today that are used in radiology, looking at breast density risk, as well as in pathology, use a number of different kinds of foundational deep learning models that have really been shown to be almost like a diagnostic for screening purposes and to estimate the risk of invasive cancer recurrence.

We are now entering a realm of a lot more text, looking at text, so that means looking at clinical notes, looking at case history, looking at different forms that we collect when a patient comes in through all their different visits from oncology. But it's not just oncology. From supportive care services to primary care and endocrinology, this is a continuum of real health encounters that we can now study, consolidate, and use AI to really help us understand the disease in a much more stratified and granular fashion.

How can these AI tools be integrated into clinical workflows without adding burden or disrupting patient care?

We don't want to have these AI tools as a separate entity, like another dashboard or another thing you have to open up and look at. These have to be integrated at the point of care and in people's—really, their everyday EHR [electronic health record] system.

There is work underway. I know big vendors, like Epic, that are doing work around this integration. There are also patient-clinician encounters that are being recorded, and so there's voice that gets converted to text, which can integrate into clinical notes.

I mean, these are all things that are happening. For these to now come together for decision-making purposes is going to require humans in the loop as well to really validate and [implement] checks and balances. These have to be embedded into the point-of-care system that patients and clinicians currently use.