Addressing care inequities in breast cancer prediction modeling
Breast cancer prediction modeling is a great example. Breast cancer screening is a traditional component of preventive care, with a growing focus on using factors like breast density to predict risk of cancer development. Black women under the age of 35 are two times more likely to be diagnosed with breast cancer than white women. Young women with breast cancer have lower survival rates than those over the age of 40 (seven percentage points). Black women are 42% more likely to die from breast cancer than white women, despite overall similar incidence rates.
Despite the urgency for action, some breast cancer measurements are unreliable and may reduce the likelihood that Black women are connected with supplemental screening. For example, even the current, most accurate model for measuring risk of breast cancer development, the Tyrer-Cuzick model (version 8), has significant flaws. The area under the curve (AUC) is a common measure of accuracy for these models. An AUC of 0.5 signifies that the model's accuracy is around 50%, whereas a perfect model would have an AUC of 1. The AUC of the Tyrer-Cuzick model is 0.62 for white women—but it's significantly worse for Black women, at just 0.45. That's worse than a coin-flip. In large part, this disparity stems from the fact that few data points from black patients were used in the model's development.
However, recent innovations in deep learning hold promise for reducing these racial disparities. These algorithms can process much larger quantities of data and include an analysis of actual mammography images.
Deep learning used to develop a better model
Researchers at Massachusetts General Hospital and MIT used the power of deep learning to develop a new, more accurate model for breast cancer risk prediction. They fed a deep learning model with full-field mammography images and a five-year outcome of whether or not the patient developed breast cancer. At the same time, the researchers created a risk-factor logistic regression model that mapped a patient's risk factors to whether or not the patient developed cancer within five years. The researchers then developed a dual model that combined both the image and the risk factor information into a single prediction of risk.
The result? A significantly better model at predicting breast cancer risk, particularly for Black women. The hybrid deep learning model is equally accurate for white and Black women with an AUC of 0.71 for both subgroups.
Determining risk from limited information
This was not the only significant result to come from the research, however, as the deep learning model based just on mammography images also outperformed the Tyrer-Cuzick model. It placed 31% of all patients with future breast cancer in the top-risk decile, compared with only 18% with the Tyrer-Cuzick model. This result means that patients who do not know all of their risk factors, such as family history, can still receive an accurate assessment of their future risk.
This new image-only model would also allow imaging screening programs to provide a more accurate assessment of risk automatically based solely on the mammogram images. This type of immediate feedback on potential risk is a large reason for the proliferation of breast density notification laws, so this new image-only deep learning model provides an opportunity for imaging screening programs to go beyond basic density notification and provide patients with a much more accurate view of their future risk.