Using A.I. to Transform Breast Cancer Care

I am very interested to know how this technique performs with extremely dense breast tissue. After six annual mammograms found no suspicious findings, I discovered an 11 mm slow-growing cancer. An MR found a second 8 mm tumor. Both tumors were “mammographically occult”; they were undetectable. Over the nine months before surgery, on treatment that should have reduced density, the radiologists said the tumors continued to remain occult. I am also very interested to know if this technique extends to interpretation of MR images. MR detects more cancers than mammograms. While the modality is very sensitive, it is not very specific. This means it reveals numerous abnormal patterns but many are false positives. Improving specificity would add immense value and spare patients from needless biopsies. (MR biopsies are particularly unpleasant.) While I agree with other readers that improvement in treatments are critically needed, we cannot begin to treat that which we cannot detect.

Maybe we need AI to tell the insurance companies that the patient needs a 3D mammogram.

Developing a more comprehensive data base on the onset and progress of breast cancer is certainly an important step forward, but my deceased wife’s recent experience of the care and protocols at Mass General that are supposedly designed to guard against or detect metastasized breast cancer, are shockingly defective. What is completely lacking is an effective theoretical model about how a breast cancer cell metastasizes. No one knows (a) why or how a single breast cancer cell breaks off from the original breast cancer tumor; (b) why the metastasized secondary cell is so much more virulent than the original breast cancer cell; and (c) how the metastasized cancer cell can travel to blood rich locations, such as the base of the brain when the original surgical evidence had suggested that no breast cancer cells had passed through the lymph system before they were surgically removed from the breast. Doctors know virtually nothing about metastasized breast cancer. This problem can only be addressed by research into the biology of metastasis (not more studies based on probability premised on the same flawed theoretical models). Two clues have recently been suggested: (1) breast cancer cells have unique chemical traits that enable them to travel through the bloodstream; and (2) errant breast cancer cells might already have traveled through the bloodstream before the original cancer cell was first detected

To collate all relevant pieces of information in cancer diagnosis, prediction of optimal care based on personal history, predisposition to cancer based on family history, analysis of its genetic and molecular make up, innumerable mutations, subtle histological staging, hormonal status, HER-2 expression to name a few is mind boggling and somehow to put all info in the mix and then expect to come out with a bespoke prescription by A.I. is a leap of fait. It will require a randomized clinical trial before its wide- spread adoption. I certainly look forward towards that day. It will take a long time to validate the results of A.I. vis a vis the traditional human approach

In 2018 I had three cancer operations. Stage four Squamous Cell Carcinoma of the tongue metastasis to the tonsils & later thyroid. The operation was done, robotically, at the Veterans Medical Center at San Diego (LaJolla). They had to cut my head half off at the neck to get all the lymph nodes out of the right side of my neck & retrieve the tumor on the base of my tongue all done ROBOTICALLY. The first operation was a tonsillectomy & this operation was worse than the actual extraction of the tumor. It just about killed me, & that is no exaggeration. The next operation was the tumor which was scarier than any of the others but later, finding out that it was the least deadly. Then when they got the labs back, they found that there was thyroid cells in some of the removed lymph nodes so then I had to have my thyroid removed & it required radiation treatment afterward to get what they could not get through the operation. I later had a CT-scan & it showed that they got all the thyroid & cancer. The really bad thing is that only 40% of people live three years after a thyroid cancer operation. Maybe I can beat the odds.

I’ve had 2 biopsies, the second revealing more abnormalities than the first and considered pre-cancerous. While I am very grateful that mammographies detected the problems early, I have been somewhat dismayed by the fact that no one was truly interested as to the cause, other than to offer me a genetic test (costly at the time) which I decided not to undergo. Dismayed because, without being a doctor, but by using reasoning, some intuition and researching online, I had deduced that it was due to ingesting ginseng, an estrogen promoter, for a short period just prior to each instance (separated by several years). Around the same time studies had shown that hormone replacement therapy (with estrogen) increased breast cancer incidence for certain women. Since I realized this, and stopped taking ginseng when feeling tired, I have been free of abnormal scans. No one bothered to ask me much about my ‘habits’ other than how long I’ve lived in Western countries (which have a higher rate of breast cancer for Asian women than those living in Asia) and since there were no prior familial cases, genetics didn’t seem to be involved. These days, I relate my self-diagnosis to my new providers when they review personal history, but I doubt the information is saved or used to inform other possible cases. They may not even believe me. Like Dr Barzilay, I’ve often had to try and heal myself by finding the root causes.

Being a breast cancer survivor (so far), I can't say enough for genetic testing. My maternal grandmother died of breast cancer in the 1960's and my mother died of ovarian cancer in 2009. I was diagnosed with stage 1, grade 3, triple negative breast cancer in 2005 after finding a lump via self breast examination, resulting in lumpectomy, chemotherapy and radiation treatments. I was diagnosed again in my other breast 9 years later with grade 0 (in situ) "suspicious" cells. Even though my surgeon knew of my history, I had to demand a genetic test. I had the BRCA gene for both breast and ovarian cancer. I immediately told my surgeon that I wanted a dual mastectomy, as well as surgery to remove all my reproductive organs (uterus, cervix, falopian tubes and ovaries). My point is that no matter what technology is out there, you're always going to have to fight with your doctors and the medical/insurance system. Mammograms and AI are well and good, and I hope it results in many more survivors, but we still have a totally messed up system in this country when it comes to actually dealing with the diagnosis, the available treatments, and worst of all: The health insurance companies.

I am a retired radiologist that did mammograms for many years. Very demanding, very frustrating and difficult. Imagine, you have 2-4 years of previous films, you have different areas of possible problems, you have to check all these areas on multiple films from many previous years, you then must decide is it worth a biopsy or not? Should i get more films or not? Should I have the patient come in in 6 months or not? Even so the return on a biopsy is 20-30 % positive. Thousands of dollars, lots of anxiety to the patient. Any thing that can improve this is welcome.

This is the most important thing happening in the entire world. Don't get distracted, what can be more important than getting closer to curing Cancer? The plethora of cancer research papers being written across the world is too much for even teams of doctors to winnow down the most relevant facts towards their particular case. If what I've been reading about Googles' progress with quantum computing comes around and quantum computing is all its cracked up to be children and cancer patients will have Hope. Please, for the sake of all Mankind, bring this Miracle to us and esp the children.

I like the use of real world data and would have appreciated comments on the false positive rates as others have noted. Kudos on the approach.

As someone who has had breast cancer, I appreciate anyone who is thinking outside the box and doing research. For those of us that had low level, stage one tumors removed, even with Onco Testing, there is always the question of whether our particular tumor is going to be the one that will metastasize.

"Internal IBM documents show that its Watson supercomputer often spit out erroneous cancer treatment advice and that company medical specialists and customers identified “multiple examples of unsafe and incorrect treatment recommendations” as IBM was promoting the product to hospitals and physicians around the world." Source: https://www.statnews.com/2018/07/25/ibm-watson-recommended-unsafe-incorrect-treatments/

No one has mentioned the fact that universal national health systems mitigate two of the constraints to more effective research mentioned here: they protect populations from social or economic harms (discrimination, etc.) associated with the possibility of disclosure of high risk health conditions, since everyone is covered, as with Medicare for those over 65, and they create large, accessible medical data bases, representative of the entire population. Using AI to obtain highly accurate data from time series mammograms is very achievable; machine reading of other imaging such as CTs and echocardiography has already been shown to be equal to !or superior in accuracy to experienced human technicians. The benefits of very early detection of breast cancer to individuals with genetic mutations that carry increased risk are obvious.

Over-hyped, once again, of newest and greatest technology. we dont really know how these cancers are driven. So doing a trillion calculations per millisecond wont change that; biologists might...

I note a variety of criticisms, e.g., methodological (what about false positives), sociological (e.g., for-profit healthcare). All worth noting. Though I would also note that a person capable of carrying out this work is well aware of measurement/methodological issues. As well, she was a "beneficiary" of less-than-optimal care and was confronted with the large amount of information that is not being harvested as data accumulates in various electronic forms. So I am as methodologically hard-nosed as the next methodologically informed person, and I work with data/analytics in Canada but also worked in healthcare in the US - so I know about issues with the systems on both sides of the border. All of that being said - this strikes ME as something worth investigating further. I want to know more about what Dr. Barzilay did and how she did it, so that I can develop/shape/refine my approaches and maybe generate something useful. One more comment re: Stephen Rinsler's comment - yes. Now if we gave more thought to how we were going to extract statistically/methodologically robust insights from our data BEFORE we architected the electronic health record documentation 'solutions', maybe we could refine our natural language processing tools on the basis of a clearer understanding of what coded data we want to extract from that natural language. And clinicians would shape their dictations to ensure appropriate interpretation. Providers and EHR's might start to get along better.

In my (more than ignorant, less than expert) view, a CRITICAL problem with managing routine clinical data is the LACK of “free text” tools. Forcing the clinician to use a data form runs the risk of eliminating much information that is capturable in narrative clinical notes. The development of a standard, open source, “capture tool” that will take down the narrative input of the clinician as she/he interacts with patients WITHOUT requiring extra work by the clinician will be a boon. Possibly current and/or soon to be released cell phones can be used with appropriate software for such data collection. Stephen Rinsler, MD

Were there false positive results? Of so what percent?

I'm a fan of Susan Gubar's writing and the way she gives a poignant voice to the cancer experience. I think it's great that Dr. Barzilay is taking a novel approach to using AI for improving detection, and I hope that she continues her work. Cancer is a such a complex disease, and there are an enormously vast number of variables that come into play and, ultimately, the modeling and generalization - garbage in is garbage out. There's no doubt that efforts like this will address some of the "low-hanging" fruit and potentially catch some cases at an earlier stage. However, I'm skeptical of the hype around AI and machine learning, and think we're only at the beginning of a very long road. The algorithms are tools, not cures.

Profit driven healthcare is what we have and AI will enhance that. What about data given to insurance companies so they can deny claims? AI might be able to transform health data diagnosis but what about the false positives and additional costs related to that? There are many questions yet to answer before AI is used extensively.

Applying A.I. to health data can transform diagnosis. In the USA system it is difficult and dangerous to share data as the insurance companies can use it to deny coverage. It begs the question, are similar studies underway in countries with better, universal health coverage such as the UK or France?

I’ve just finished 6 rounds of chemo to fight breast cancer and have been amazed that this sort of data analysis is not already employed to analyze breast images and related data. I’ve been a vegetarian for 25 years, am lean, love to exercise, non smoker, very light drinker, live in a relatively clean city, clean my house with baking soda, no genetic propensity to breast cancer. Yet with dense breast tissue (like 40% of US women) and a 2015 mammogram that required ultrasound imaging of my right breast where my cancer now grows, I have been at the mercy of HealthPartners insurance company decision making. My nurse midwife wrote a referral for a 3D mammogram for me last year, but HealthPartners denied it; my resulting 2D mammogram came back clear. In 2019 data trickled in from a 3D mammogram, biopsies, MRI, finally revealing three cancer masses. If Barzilay’s data analysis could have been utilized, the tiny changes in my 6 years of mammograms (I’m 46) could have been spotted; I could have been flagged as someone who “deserved” access to the 3D mammogram. Since my cancer was diagnosed later, I’ve had to endure 6 cycles of chemo and its many permanent effects, had a port installed in my chest resulting in deep vein thrombosis in my jugular and now 9 months on blood thinner, mastectomy surgery next month, at least two more surgeries, possible radiation, and chemo with cancer-targeting drugs for 7 more months. Sure prevention is ideal and yet early detection is vital!

This is amazing! Having just negotiated the medical system for my mother's breast cancer, accurate information about what worked for whom would have been appreciated. Moreover, anything that would help catch this disease in its early stages or predict it with more accuracy would be so helpful. The commenters complaining about wanting a cure instead are missing the point. Once computers know enough about who gets cancer and when and where they live and what they eat and how they move, etc, they will be able to predict what, exactly is causing the cancer. Cancer is a metabolic disease and we should ultimately be able to prevent it, which is way better than curing it.

What Dr. Barzilay is describing is a capitalistic problem and not a scientific or computational problem. We have a profit driven health care system and not a disease curing or preventative health care system. The profit driven companies that operate these hospitals and cancer centers only want one outcome. More $$$ and not more cures. What good does developing a computer system that can better detect early cancer signs if greedy CEOs price the use of it so high that it forces people to decide on bankruptcy or getting early detection?

Yes, a “travesty.” In my experience, there are so many aspects of medicine that are art, not data-driven science. And yet that is not how physicians present information or recommendations to patients. There is much work in this area to be done.

This article begs more questions than it answers. For one, if the AI was programmed based on a variety of mammograms, it may be able to detect breast cancer earlier, but I fail to see how it will be able to determine treatment. Even if it's programmed with numerous pathology reports, at best, it will only be able to make treatment recommendations based on the 3 classic subtypes (estrogen postive, HER2 postive or triple negative) which are already available to patients once they're diagnosed. I fail to see how it would open up clinical trials for people nor do I see how it will open up new treatment options that are now available because of genomic testing of tumors or genetic testing of patients. At best, it may improve screening and earlier detection and even there, research has shown that metastasis can happen long before a tumor shows up in the breast. There's still a long way to go before there's a cure or prevention model.

"A quest to democratize cancer care." It's about time. When will this "research in progress" be utilized in real time?

Although the author wrote that “computers can be employed to map the medical history of many patients so as to predict and better treat new cancers — and maybe to prevent them,” I didn’t see anything in the article about actual prevention of cancer. To me, it looked like earlier detection of cancer, which can then be treated. But by the time breast cancer can be detected on a mammogram, it has usually been growing for years; I’ve read estimates of up to 20 years. And during most of those years, the tumor is shedding cancer cells. So, even detecting it a few years earlier does not prevent it. I would love to see much more research on how to actually prevent breast cancer. So far, there are five risk factors: excess body weight, poor diet, lack of exercise, smoking tobacco, and drinking alcohol. Improving any of these risk factors decreases the risk of breast cancer; improving them all results in the biggest decrease in risk. And these factors are all under our control. But the risk does not go to zero. Perhaps because we breathe contaminated air and drink polluted water. But these factors are out of our control, unless we work with our government officials to increase and enforce regulation of polluters, and with industry to decrease pollution.

As a 2 time traveler with this horrible disease I think the technology shows promise. That said, I am most worried that the insurance companies and employers will use this information against people who either had breast cancer or might be prone to that. If that is the case, I don’t believe the benefits will outweigh the risks and that the data will be used first to hurt women.