The company trained its AI using de-identified data from patients in both the US and the UK, and showed that it could reduce false positives by 5.7 percent and false negatives by 9.4 percent in the US. Interestingly, a smaller reduction of 1.2 percent and 2.7 percent respectively was seen in the UK, suggesting that the current US detection system has lower accuracy than the current UK system.

Unlike the human experts, who used patient histories and prior mammograms to make their assessments, the AI only had access to the most recent mammogram of each patient. Despite this, it was able to make screening decisions with greater accuracy than the experts, and the model could be generalized to different populations — such as women in the US compared to women in the UK.

The developers of the AI emphasize that this is early stage research and that more studies and cooperation with healthcare providers will be required before the system is ready for widespread use.

DeepMind has been used in the past for medical purposes from spotting eye diseases to predicting kidney illness, however, it has also been the subject of considerable controversy. In 2017, it was revealed that the UK’s National Health System had shared data with DeepMind on an “inappropriate legal basis,” with the company receiving 1.6 million patient records without the direct consent of the patients. This broke privacy laws, the UK data watchdog ruled, so the NHS chose continue working with DeepMind but to anonymize data in future.

In 2018, DeepMind was brought under the Google Health initiative, and concerns about privacy were not assuaged when Google dissolved the review board which was supposed to oversee the company’s relationship with the NHS. For all the potential good that could be done with a medical AI like DeepMind, there seems to be a concerning lack of oversight over the privacy of patient data and a lack of accountability for past data privacy issues.

The IBM approach trains the AI with anonymized mammography images linked to biomarkers (such as reproductive history) and clinical data, allowing the creation of an algorithm with comparatively high accuracy. It can reduce the chance of a bad diagnosis by establishing connections between traits you wouldn’t spot in imagery alone, such as iron deficiencies and thyroid function. IBM even pulls in data from biopsies, lab tests, cancer registries and codes from other diagnoses and procedures.

You wouldn’t want to rely solely on the algorithm to make predictions, especially when it correctly interprets just 77 percent of non-cancerous instances. However, the accuracy is good enough that it could serve as a “second set of eyes,” according to IBM. It could verify a radiologist’s prognosis and reduce the chances of patients being sent in for unnecessary follow-up tests. This could be particularly important in countries where staff shortages make it impractical for another radiologist to weigh in, or in any situation where there isn’t much time for human checks.

This isn’t always going to be the most advanced form of breast cancer prediction. MIT’s recently developed method works up to five years in advance using just images. However, IBM is betting that a more holistic approach will be more worthwhile to doctors, and its approach may be more reflective of the overall population simply by looking at non-image factors that are common to everyone. Either way, there’s a real possibility that more breast cancer patients will start treatment before the first tumor even appears.

The scientists first looked at the mammograms of over 60,000 patients who were treated at Massachusetts General. They then identified the women that developed breast cancer within five years of their screening. With this data, scientists created a model that recognizes the subtle patterns in breast tissue that are the early signs of cancer. The results of the study can be found in a paper published this week in the journal Radiology.

AI has potential to help fix the racial disparity in women’s healthcare as well. Since current guidelines for breast cancer are based on primarily white populations, this can lead to delayed detection among women of color, a 2018 report in JAMA Surgery found. This has lead to severe consequences; Black women are 43 percent more likely to die from breast cancer than white women. On average, Hispanic, black and Asian women develop breast cancer at an earlier age than their white counterparts.

Scientists found that their AI model worked on both black patients and white patients for a simple reason; their training data included both populations. “It’s particularly striking that the model performs equally as well for black and white people, which has not been the case with prior risk assessment tools”, said Dr. Allison Kurian, associate professor of Medicine and Health Research and Policy at Stanford University to MIT. “If validated and made available for widespread use, this could really improve on our current strategies to estimate risk.”