Mammograms are currently the gold standard for breast cancer screening, but there is controversy over when and how often to administer them. On the one hand, advocates claim their ability to save lives. For example, a 60-69 year-old woman with a mammogram has a 33% lower risk of dying than a woman without a mammogram. On the other hand, some people discuss costly and potentially traumatic false positives. A meta-analysis of three randomized trials found a 19% overdiagnosis rate on mammography.
Even with some lives saved and some overtreatment and overscreening, the current guidelines are still catch-all. Women aged 45-54 need to get a mammogram every year. Personalized screening has long been the answer, but tools that can leverage the wealth of data to do so are lagging behind.
With this, scientists at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Jameel Clinic for Machine Learning and Health asked: Can you use machine learning to provide personalized screening?
This gave rise to Tempo, a technology for creating risk-based screening guidelines. Using an AI-based risk model that looks at who was screened and when the diagnosis was made, Tempo recommends that patients return to mammography at specific times in the future such as 6 months or 3 years. The same Tempo policy can be easily applied to a variety of possible screening options, allowing clinicians to choose the desired compromise between early detection and screening costs without learning new policies.
The model was trained on a large screening mammography dataset from Massachusetts General Hospital (MGH) and tested on MGH patients and external datasets from Emory, Karolinska Sweden and Chang Gung Memorial Hospitals. Using the Mirai risk-assessment algorithm the team previously developed, Tempo achieved better early detection than annual checkups, and at Karolinska Hospital required 25% fewer mammograms. MGH recommended mammography for about a year and improved the modeled benefit of early detection by about four and a half months.
“By tailoring screening to the individual risk of the patient, we can improve patient outcomes, reduce overtreatment and eliminate health disparities,” said a graduate student in electrical engineering and computer science. A member of MIT CSAIL, Pace describes a paper published in Nature Medicine on January 13. “Given the scale of breast cancer screening, where tens of millions of women undergo mammography each year, improving guidelines is very important.”
The earliest uses of AI in medicine date back to the 1960s, when many cited as the beginnings of Dendral’s experiments. Researchers have created a software system that organic chemists consider to be the first types of experts to automate decision-making and problem-solving. Sixty years later, deep medicine has made significant advances in drug diagnosis, predictive medicine, and patient care.
“Current guidelines divide the population into several broader groups, such as groups under 55 or over 55, and recommend the same frequency of screening for all members of the cohort. By developing an AI-based risk model that processes patient data, we have the opportunity to transform screening, screen people who need it more often, and save the rest, “says Yala. I am. “An important aspect of these models is that predictions can change over time as the patient’s raw data changes, which adjusts screening guidelines to change in risk and for long-term patients. It suggests that it needs to be optimized over long periods of patient data.”
Tempo uses reinforcement learning, a machine learning technique widely known for success in games like chess and Go, to develop “policy” that predicts the next recommendation for each patient.
The training data here contained only information about the patient’s risk when the mammogram was taken (for example, 50 or 55 years old). The team needed a mid-point risk assessment, so they designed an algorithm to learn patient risk at unobserved time from observed screening that evolved as new patient mammograms became available. ..
The team first trained the neural network to predict future risk assessments based on previous risk assessments. The model then estimates the patient’s risk at an unobserved time point and allows simulation of risk-based screening guidelines. They then trained this guideline (which is also a neural network) to maximize rewards for retroactive training sets (e.g., a combination of early detection and screening costs). You will eventually receive recommendations on when to return to the next screen ranging from 6 months to 3 years. Only 1-2 years is standard.
Let’s say Patient A comes to the first mammogram and is finally diagnosed in the 4th year. They have nothing in the second year, so they don’t come back for another two years, but they get diagnosed in the fourth year. Currently, there is a two-year gap between the last screens where the tumor may have grown.
With the Tempo, the first mammogram of the year zero may have recommended returning after two years. And in the second year, the risk turns out to be high, recommending that the patient return within 6 months, and at best be able to detect it. The model dynamically changes the patient’s screening frequency as the risk profile changes.
Tempo uses a simple early detection indicator that suggests cancer can be detected within 18 months. Tempo outperformed current guidelines in various conditions (6 months, 12 months) of this assumption, but neither of these assumptions is perfect, as the likelihood of early detection of a tumor depends on the characteristics of that tumor. The team suggested that follow-up work with tumor growth models could address this issue.
In addition, the Tempo-recommended screening cost metric, which counts the total amount of screening, does not explicitly quantify the risk of false positives or the harm of additional screening, and therefore does not provide a complete analysis of future total cost.
There are many future directions in which personalized screening algorithms can be further improved. The team states that one method is to build on the metrics used to estimate the cost of early detection and screening from retroactive data, which leads to more sophisticated guidelines. Tempo can also be adjusted to accommodate different types of screening recommendations, including: For example, with the use of MRI or mammograms, and future work, the costs and benefits of each can be modeled individually. With better screening guidelines, it may be possible to recalculate the earliest and latest age that are cost-effective for patients to be screened.
“Our framework is flexible and easy to use for other diseases, other forms of risk models, and other definitions of early detection benefits and screening costs. As risk models and outcome measures continue to improve, we expect the usefulness of Tempo to continue to improve. We are excited to work with our hospital partners to advance this technology and further improve personalized cancer screening,” says Yala.
