SABCS 2024: Identifying High-Risk Women for Targeted Breast Cancer Prevention Through AI-Powered Mammography Analysis

In an interview with Pharmacy Times®, Mikael Eriksson, PhD, epidemiologist at Karolinska Institute in Sweden, shared the aim behind the use artificial intelligence (AI)-based analysis of mammography images to assess long-term breast cancer risk, to identify women who could benefit from preventive interventions. This long-term risk assessment is distinct from using AI to identify short-term risk for improved screening. Eriksson noted that implementing an AI-based risk model in clinical practice requires overcoming challenges like obtaining diverse training data and navigating regulatory approval.

Pharmacy Times

How can AI-powered image analysis be used to identify women at high risk of developing breast cancer?

Mikael Eriksson, PhD

There are two different clinical aspects that we are interested in. One is to identify a risk in the short term, 1 to 5 years, to identify women who benefit from additional imaging. That is to improve the screening outcomes. But what we have done in this study is to look into risk assessment in the long term, that is to identify women who may benefit from a risk reducing intervention. We can then offer women at study baseline lifestyle intervention or medical intervention, and that is for reducing the risk of breast cancer and reducing incidence of breast cancer.

Pharmacy Times

What are the potential benefits of using an AI risk model for primary breast cancer prevention, such as early detection and targeted interventions?

Eriksson

The most important aspect that we're showing here in this study is that we are assessing long term risk, and that is specifically for identifying women where we can actually prevent the women from developing a breast cancer. For that reason, we assess risk in the long term, 10 years because it takes a long time for tumor to progress into a clinically identifiable cancer. So, this aspect is the most important for primary prevention, but then we have the other aspects of identifying risk in the short term, and that is for improving screening outcomes. They have two different clinical targets, and we are developing models for both aspects. But in this specific abstract we are talking about the long promise primary prevention.

Pharmacy Times

What are the challenges in developing and implementing an AI-based risk model in clinical practice, and how can these challenges be addressed?

Eriksson

We are using image data, which is really an advantage, because we are utilizing the mammography screening units. We have a lot of images generated that, but we need to establish collaborations with a lot of clinics, because we need to have a big variety of data. We know that AI learning very well, but it can also be biased in that sense that they start to learn specific vendors, specific screening routines and so on. We need to have a big mix of data in order to do that.

The second question about how we can implement this in the clinic — we need to go through regulatory pathways. It has been now decided by FDA that we need to have FDA approval for doing this risk model. That is a big workload, and the way that we can handle that is to even develop bigger cohorts, so we can both train and validate the cohort in very various data sets, so it could be proof of clinical use.