A team of researchers from MIT and Massachusetts General Hospital (MGH) have created an AI model capable of predicting a patient’s breast cancer risk five years in advance.
The deep learning (DL) model was also equally as accurate for racial minorities who have proven to be more likely to die from cancer, such as black women, according to a May 7 study published in Radiology.
“It’s particularly striking that the model performs equally as well for white and black people, which has not been the case with prior tools,” said Allison Kurian, an associate professor of medicine and health research/policy at Stanford University School of Medicine, in an email news release. “If validated and made available for widespread use, this could really improve on our current strategies to estimate risk.”
The retrospective study included 88,994 consecutive screening mammograms taken from nearly 40,000 women. All exams were broken into a training set (71,689), validation group (8,554) and a test set of 8,751.
Lead author Adam Yala, a PhD student at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and colleagues used patient questionnaires and electronic medical records to develop three AI models: a risk-factor-based logistic regression model (RF-LR) based on traditional risk factors, a DL model (image-only DL) that only used mammograms and a hybrid model which combined the two.
The DL models were compared to the Tyrer-Cuzick model, version 8—an established breast cancer risk model that includes breast density.
Overall, the hybrid model placed 31% of all patients with future breast cancer in the top-risk category compared to the 18% placed using the Tyrer-Cuzick model. In the test set, the latter had an area under the receiver operating characteristic curve (AUC) of 0.62, while the hybrid model notched a 0.70, RF-LR scored a 0.67 and image-only DL scored a 0.68.
The hybrid model was also equally accurate for white and African American Women, with an AUC of 0.71 for both ethnicities, compared to the Tyrer-Cuzick model which resulted in an AUC of 0.62 and 0.45 for white and black women, respectively.
According to the researchers, the hybrid model still suffers from the black box problem seen in many AI methods, but they speculated that it might use different fine-grain tissue patterns and “relative orientations of those patterns depending on global patterns in a patient’s breast,” which identify distinguishing patterns for women with dense and nondense breasts.
More research is needed, included a validation study performed across multiple institutions and vendors, before the model is ready for use, but the models may be able to replace traditional risk factors in determining breast cancer risk and could eventually enable a more individualized approach to breast cancer care.
“Since the 1960s radiologists have noticed that women have unique and widely variable patterns of breast tissue visible on the mammogram,” said co-author Constance Lehman with MGH, in the same release. “These patterns can represent the influence of genetics, hormones, pregnancy, lactation, diet, weight loss and weight gain. We can now leverage this detailed information to be more precise in our risk assessment at the individual level.”
The authors disclosed that MIT and MGH have patents filed on the deep learning models.