The COVID-19 pandemic puts routine health screenings on hold, making it more likely for cancer to go undetected.
Constance Lehman (MD, PHD, Professor of Radiology at Harvard Medical School) designs machine-learning models that aid in breast cancer detection, which became critical tools as screenings went on hold during COVID-19 lockdowns. As Chief of Breast Imaging at Massachusetts General Hospital, Lehman collaborates with Regina Barzilay (MIT Computer Science & Artificial Intelligence Lab) from Jameel Clinic on this project. Lehman will be speaking about "Clinical Implementation of AI Tools to Reduce Disparities and Increase Access in Peri-pandemic Times" at the AI Cures Conference on September 29th.
In March, Massachusetts Governor Charlie Baker issued an emergency order that halted appointments and surgeries that were not urgent, including mammograms to screen for breast cancer, so hospitals could divert resources to COVID-19 patients. By the time screening appointments reopened in May, MGH had cancelled mammograms for about 12,000 women.
What happened in Massachusetts reflects the United States at large, Lehman says. Normally, clinics across the states conduct nearly 100,000 mammograms in the US each day. While screenings were postponed at different times depending on the state, millions of women have forgone their routine screening since March.
Now, hospitals are focusing on gradually rescheduling these patients. To ease the process, MGH is assessing patient risk with a machine-learning model developed by Drs. Barzilay and Lehman. The model helps identify patients who should be invited back first by predicting who is likely to develop breast cancer in the next five years.
Traditional risk models rely on factors like breast density, reproductive and family history, but they are not refined enough to determine which patients should be prioritized for screening. “The vast majority of women diagnosed with breast cancer do not have any known risk factors, except that they are female and they are over 40,” Lehman says. Traditional models have failed to accurately identify individual women at increased risk, and also significantly underperform for people of color compared to white patients.
In contrast, the MGH-MIT machine-learning model picks up on complex patterns in the images that humans would otherwise be unable to see.