While mammograms are currently the gold standard in breast cancer screening, whirlpools of controversy exist about when and how often they should be administered. On the one hand, advocates argue for life-saving potential: Women aged 60-69 who get mammograms, for example, have a 33 percent lower risk of dying than those who don’t. Meanwhile, others argue about costly and potentially painful false positives: A meta-analysis of three randomized trials found a 19 percent higher diagnosis rate than mammography.
Even with some lives saved, and with some overtreatment and overscreening, current guidelines are still a catch-all: Women between the ages of 45 and 54 should get mammograms every year. While individual screening has long been regarded as the answer, tools that can take advantage of the trove of data to make it so lag behind.
It asked scientists at MIT’s Computer Science and Artificial Intelligence Laboratory (CSEL) and the Jameel Clinic for Machine Learning and Health: Can we use machine learning to provide personalized screening?
Out of this came Tempo, a technique for creating risk-based screening guidelines. Using an AI-based risk model that looks at who was screened and when they were diagnosed, Tempo will recommend a patient return for a mammogram at a specific time point in the future, such as six months. or three years. The same tempo policy can be easily adapted to a wide range of potential screening preferences, which will let clinicians choose their desired early-detection-to-screening-cost trade-offs, without the need for training in new policies.
The model was trained on a large screening mammography dataset from Massachusetts General Hospital (MGH), and tested on patients outside of MGH as well as external datasets from Emory, Karolinska Sweden and Chang Gung Memorial Hospitals. Using Mirai, the team’s previously developed risk-assessment algorithm, Tempo achieved better early detection than annual screening, while Karolinska required 25 percent fewer mammograms overall. At MGH, it recommended a mammogram at roughly a year, and a spurious early detection benefit of about four and a half months better.
“By tailoring screening to individual patient risk, we can improve patient outcomes, reduce over-treatment, and eliminate health disparities,” said Adam Yala, Electrical Engineering and Computer Science PhD student in the U.S., affiliated with MIT CSAL, and lead researcher on a paper, says. The description of Tempo Published on January 13 nature medicine, “Given the massive scale of breast cancer screening, with millions of women getting mammograms each year, improving our guidelines is extremely important.”
The early use of AI in medicine dates back to the 1960s, where many refer to dendral experiments as getting off the ground. Researchers created a software system thought to be the first specialist type that automates the decision-making and problem-solving behavior of organic chemists. Sixty years later, Deep Medicine has developed greatly in drug diagnostics, predictive medicine, and patient care.
“Current guidelines divide the population into somewhat larger groups, such as those under or over 55 years of age, and recommend equal screening frequency for all members of a group. Development of an AI-based risk model that uses raw patient data works on, gives us the opportunity to replace screening with those who need it and leave the rest to screen more often,” Yala says. “A key aspect of these models is that their predictions can evolve over time as a patient’s raw data changes, suggesting that screening policies should be correlated with changes in risk and the need for patient data.” Must be optimized over the long term.”
Tempo uses reinforcement learning, a machine learning method widely known for its success in games such as chess and Go, to develop a “policy” that predicts a follow-up recommendation for each patient.
The training data here only contained information about the patient’s risk at the time their mammogram was taken (when they were 50 or 55 years old, for example). The team needed risk assessment at intermediate points, so they designed their algorithm to learn a patient’s risk at unobserved time points from their observed screening, which evolved as new patient mammograms became available.
The team first trained a neural network to predict future risk assessments over the past. This model then estimates patient risk at undiscovered time points, and it enables the simulation of risk-based screening policies. Next, they trained that policy, (also a neural network) to maximize reward in a retrospective training set (for example, a combination of early detection and screening cost). Eventually, you’ll get a recommendation for when to return to the next screen, from six months to three years in the future, in multiples of six months—the standard is only a year or two.
Let’s say patient A comes in for her first mammogram, and is eventually diagnosed in her fourth year. In year two, there is nothing, so they don’t come back for two more years, but then in year four they get a diagnosis. There is now a gap of two years between the last screen where the tumor could have been.
Using Tempo, at that first mammogram, year zero, the recommendation may be to return in two years. And then in the second year, it can be seen that the risk is high, and recommended that the patient come back in six months, and in the best case scenario, it will be detectable. The model is dynamically changing the patient’s screening frequency based on how the risk profile is changing.
Tempo uses a simple metric for early detection, which assumes cancer can be caught up to 18 months earlier. While TEMPO outperformed existing guidelines on this assumption (six months, 12 months) in various settings, neither of these assumptions is correct, as the early detection ability of a tumor depends on the characteristics of that tumor. The team suggested that follow-up work using a tumor growth model could address this issue.
In addition, the screening-cost metric, which calculates the total screening volume recommended by Tempo, does not provide a complete analysis of future overall costs because it does not clearly quantify false positive risks or additional screening losses.
There are several future directions that could further improve personalized screening algorithms. The team says that building on the metrics used to estimate early detection and screening costs from retrospective data will result in more sophisticated guidelines. Tempo can also be adapted to incorporate a variety of screening recommendations, such as taking advantage of MRI or mammograms, and future work may model the costs and benefits of each separately. With better screening policies, it may be possible to recalculate the earliest and latest age that screening is still cost-effective for a patient.
“Our framework is flexible and can be readily used for other diseases, other forms of risk models, and other definitions of early detection benefit or screening cost. We expect the usefulness of Tempo to continue to improve as risk models And as outcome metrics are further refined. We are excited to potentially study this technology and work with hospital partners to help further improve personalized cancer screening,” says Yala.
Yala co-authored the paper on tempo with MIT PhD student Peter G. Mikhail, Fredrik Strand of Karolinska University Hospital, Gigin Lin of Chang Gung Memorial Hospital, Yung-Liang Wan of Chang Gung University, Siddhartha Satuluru of Emory University, Thomas Kim of Georgia Tech . , Hari Trivedi of Emory University, Eamon Banerjee of the Mayo Clinic, Judy Gichoya of Emory University School of Medicine, Kevin Hughes of MGH, Constance Lehmann of MGH, and senior author and MIT professor Regina Barzile.
Research by Susan G. Komen, Breast Cancer Research Foundation, Quanta Computing, an Anonymous Foundation, MIT Jamil-Clinic, Chang Gung Medical Foundation Grant and Stockholm Lance Landsting HMT Grant.
