WHEN you are diagnosed with cancer, you are faced with a lot of uncertainty and forced to make dramatic decisions often based on very little data. That鈥檚 what found when she was told she had breast cancer in 2014. But data is her bread and butter 鈥� she works on machine learning, teaching computers to read language or predict outcomes based on a few clues. And with cancer, that鈥檚 all we have right now: the clinical data on which doctors base a patient鈥檚 prognosis is drawn from just a sliver of the population. Barzilay wants to change that.
A professor who teaches one of the most popular classes at the Massachusetts Institute of Technology 鈥� an introduction to machine learning 鈥� and a , Barzilay is part of a revolution brewing in cancer detection. The team she leads at MIT is using artificial intelligence to recognise patterns in medical images and doctors鈥� electronic notes in an attempt to catch cancer earlier and avoid overusing invasive treatments.
How did you end up applying your work with machine learning to oncology?
Advertisement
There is a technical answer and a pragmatic one. Virtually every aspect of life today is regulated by machine learning, whether you know it or not. The only area that isn鈥檛 is healthcare, which involves a lot of prediction tasks. When your doctor tries to find you a treatment, they look at different clues together and make a prediction. With personalisation, which we鈥檙e all trying to achieve in medicine, the goal is matching you and your unique characteristics to the correct drug. I got into oncology because I was a breast cancer patient at Massachusetts General Hospital (MGH).
Why do we need machine learning for personalised medicine?
Even very experienced doctors have only seen a limited number of patients. Maybe none of them were exactly like you. That鈥檚 particularly true for cancer, where there is so much trial and error. A human can look at a scan and summarise what they see, but when they sum up an image that has hundreds of thousands of pixels as a single page of words, all the unique information is lost. We鈥檙e in the 19th century here, in science terms. There was so much opportunity and possibility that after I finished my cancer treatment, I came back to MIT and I started bringing what we do there to MGH and to oncology more generally.
Was there a particular moment during your treatment that led you to want to work on this?
It was at every single step. For instance, after my initial diagnosis, I had a mammogram and they said this is a very small cancer. Then they did an MRI, which they do before surgery, and the cancer was all over the place. So they needed to do a biopsy, which found that it was a false positive. The cancer was not everywhere. Why don鈥檛 we train a machine to predict what鈥檚 going on, instead of doing these painful, expensive procedures? Every single step of the way, the reason we selected a certain treatment for me was that we weren鈥檛 sure. We were just going for the most aggressive thing.
鈥淢illions get breast cancer, but decisions are based on just 3 per cent of them鈥�
After the diagnosis, I started reading about my prognosis. For instance, after the initial treatment, there was consideration to giving me a drug with terrible side effects. There was a big clinical study involving the drug in The New England Journal of Medicine and I thought 鈥渢hat鈥檒l tell us what to do鈥�. But even though it was a big study, only a small subset of women had been in my specific situation.
Learn more at New Scientist Live:
Millions of women get breast cancer, but I found out later that all the oncology decisions made in the US today are based on the 3 per cent of those women who participate in clinical trials. That鈥檚 very problematic: what is the chance that the women in that 3 per cent will be like you or me? But if you look at the whole population of women with breast cancer, the chances are significantly higher.
What are the privacy concerns with gathering people鈥檚 health data in a massive set like this?
We think about that from the design stage onwards. To get access to any data, we need approval from an institutional review board at the hospital to make sure it is handled according to protocol. The data lives at the hospital 鈥� we don鈥檛 bring it to MIT. We also anonymise it. Any workable system needs to be totally integrated with clinical care, observing the hospital鈥檚 rules of patient privacy and rights.
What about bad data? Can that bias your machine-learning model?
Certainly. If the hospital鈥檚 pathologists miss cancer diagnoses or misdiagnose something else as cancer, the system will be trained on noisy data. We need to make sure that whatever human-generated information we are using t is as clean as possible.
Another type of bias we have seen is one where a certain population is under-represented in the general patient population. That leads to higher error rates. In the hospitals we are generally working with, there has been a smaller proportion of African-American women. We can address this algorithmically, but we are also collecting data from hospitals that have a reasonable representation of diverse groups.
If we apply machine learning to diagnosis or treatment, is there still a role for the doctor?
Absolutely. In some cases, the machine rivals the human. Sometimes it performs below the human. But the real power comes when you put the two together. Still, the doctor doesn鈥檛 have to agree with the machine. Ultimately, the doctor makes the decisions and approves treatment.
How are you applying machine learning to the problem of cancer overdiagnosis?
We are reducing the uncertainty. There鈥檚 a condition found in breasts called high-risk lesions. In the US, all patients with these lesions get surgery. But 87 per cent of them are in fact benign: the patients didn鈥檛 need the operation. With machine learning, we can now identify 30 per cent of the women who don鈥檛 need the surgery. That鈥檚 not all of them, but it is a big step.
If your system had been in use at the hospital when you were a patient, would you have felt reassured?
Oh, absolutely. I was diagnosed in 2014. But if you look again at my mammogram from 2013, you already see cancer there. When you look at the one from 2012, you already see some small spots. Maybe we didn鈥檛 know what it was. How would my life have been different if I had been diagnosed in 2012? Maybe I didn鈥檛 need to lose my hair and all the other things that came with my treatment. As a patient, not a scientist, I think we have to do everything to make sure all this uncertainty is resolved and we are applying the best technologies available to what we care about the most 鈥� our health.
This article appeared in print under the headline 鈥淚 teach machines to hunt down cancer鈥�
