The Algorithm Will See You Now
Artificial intelligence in medicine, still in an early phase, stands to transform how doctors and nurses spend their time.
There's a quiet revolution going on in medicine. It's driven by artificial intelligence, but paradoxically, new technology may put a more human face on healthcare.
AI's usefulness in healthcare ranges far and wide.
Artificial intelligence is software that can process massive amounts of information and learn over time, arriving at decisions with striking accuracy and efficiency. It offers greater accuracy in diagnosis, exponentially faster genome sequencing, the mining of medical literature and patient records at breathtaking speed, a dramatic reduction in administrative bureaucracy, personalized medicine, and even the democratization of healthcare.
The algorithms that bring these advantages won't replace doctors; rather, by offloading some of the most time-consuming tasks in healthcare, providers will be able to focus on personal interactions with patients—listening, empathizing, educating and generally putting the care back in healthcare. The relationship can focus on the alleviation of suffering, both the physical and emotional kind.
Challenges of Getting AI Up and Running
The AI revolution, still in its early phase in medicine, is already spurring some amazing advances, despite the fact that some experts say it has been overhyped. IBM's Watson Health program is a case in point. IBM capitalized on Watson's ability to process natural language by designing algorithms that devour data like medical articles and analyze images like MRIs and medical slides. The algorithms help diagnose diseases and recommend treatment strategies.
But Technology Review reported that a heavily hyped partnership with the MD Anderson Cancer Center in Houston fell apart in 2017 because of a lack of data in the proper format. The data existed, just not in a way that the voraciously data-hungry AI could use to train itself.
The hiccup certainly hasn't dampened the enthusiasm for medical AI among other tech giants, including Google and Apple, both of which have invested billions in their own healthcare projects. At this point, the main challenge is the need for algorithms to interpret a huge diversity of data mined from medical records. This can include everything from CT scans, MRIs, electrocardiograms, x-rays, and medical slides, to millions of pages of medical literature, physician's notes, and patient histories. It can even include data from implantables and wearables such as the Apple Watch and blood sugar monitors.
None of this information is in anything resembling a standard format across and even within hospitals, clinics, and diagnostic centers. Once the algorithms are trained, however, they can crunch massive amounts of data at blinding speed, with an accuracy that matches and sometimes even exceeds that of highly experienced doctors.
Genome sequencing, for example, took years to accomplish as recently as the early 2000s. The Human Genome Project, the first sequencing of the human genome, was an international effort that took 13 years to complete. In April of this year, Rady Children's Institute for Genomic Medicine in San Diego used an AI-powered genome sequencing algorithm to diagnose rare genetic diseases in infants in about 20 hours, according to ScienceDaily.
"Patient care will always begin and end with the doctor."
Dr. Stephen Kingsmore, the lead author of an article published in Science Translational Medicine, emphasized that even though the algorithm helped guide the treatment strategies of neonatal intensive care physicians, the doctor was still an indispensable link in the chain. "Some people call this artificial intelligence, we call it augmented intelligence," he says. "Patient care will always begin and end with the doctor."
One existing trend is helping to supply a great amount of valuable data to algorithms—the electronic health record. Initially blamed for exacerbating the already crushing workload of many physicians, the EHR is emerging as a boon for algorithms because it consolidates all of a patient's data in one record.
Examples of AI in Action Around the Globe
If you're a parent who has ever taken a child to the doctor with flulike symptoms, you know the anxiety of wondering if the symptoms signal something serious. Kang Zhang, M.D., Ph.D., the founding director of the Institute for Genomic Medicine at the University of California at San Diego, and colleagues developed an AI natural language processing model that used deep learning to analyze the EHRs of 1.3 million pediatric visits to a clinic in Guanzhou, China.
The AI identified common childhood diseases with about the same accuracy as human doctors, and it was even able to split the diagnoses into two categories—common conditions such as flu, and serious, life-threatening conditions like meningitis. Zhang has emphasized that the algorithm didn't replace the human doctor, but it did streamline the diagnostic process and could be used in a triage capacity when emergency room personnel need to prioritize the seriously ill over those suffering from common, less dangerous ailments.
AI's usefulness in healthcare ranges far and wide. In Uganda and several other African nations, AI is bringing modern diagnostics to remote villages that have no access to traditional technologies such as x-rays. The New York Times recently reported that there, doctors are using a pocket-sized, hand-held ultrasound machine that works in concert with a cell phone to image and diagnose everything from pneumonia (a common killer of children) to cancerous tumors.
The beauty of the highly portable, battery-powered device is that ultrasound images can be uploaded on computers so that physicians anywhere in the world can review them and weigh in with their advice. And the images are instantly incorporated into the patient's EHR.
Jonathan Rothberg, the founder of Butterfly Network, the Connecticut company that makes the device, told The New York Times that "Two thirds of the world's population gets no imaging at all. When you put something on a chip, the price goes down and you democratize it." The Butterfly ultrasound machine, which sells for $2,000, promises to be a game-changer in remote areas of Africa, South America, and Asia, as well as at the bedsides of patients in developed countries.
AI algorithms are rapidly emerging in healthcare across the U.S. and the world. China has become a major international player, set to surpass the U.S. this year in AI capital investment, the translation of AI research into marketable products, and even the number of often-cited research papers on AI. So far the U.S. is still the leader, but some experts describe the relationship between the U.S. and China as an AI cold war.
"The future of machine learning isn't sentient killer robots. It's longer human lives."
The U.S. Food and Drug Administration expanded its approval of medical algorithms from two in all of 2017 to about two per month throughout 2018. One of the first fields to be impacted is ophthalmology.
One algorithm, developed by the British AI company DeepMind (owned by Alphabet, the parent company of Google), instantly scans patients' retinas and is able to diagnose diabetic retinopathy without needing an ophthalmologist to interpret the scans. This means diabetics can get the test every year from their family physician without having to see a specialist. The Financial Times reported in March that the technology is now being used in clinics throughout Europe.
In Copenhagen, emergency service dispatchers are using a new voice-processing AI called Corti to analyze the conversations in emergency phone calls. The algorithm analyzes the verbal cues of callers, searches its huge database of medical information, and provides dispatchers with onscreen diagnostic information. Freddy Lippert, the CEO of EMS Copenhagen, notes that the algorithm has already saved lives by expediting accurate diagnoses in high-pressure situations where time is of the essence.
Researchers at the University of Nottingham in the UK have even developed a deep learning algorithm that predicts death more accurately than human clinicians. The algorithm incorporates data from a huge range of factors in a chronically ill population, including how many fruits and vegetables a patient eats on a daily basis. Dr. Stephen Weng, lead author of the study, published in PLOS ONE, said in a press release, "We found machine learning algorithms were significantly more accurate in predicting death than the standard prediction models developed by a human expert."
New digital technologies are allowing patients to participate in their healthcare as never before. A feature of the new Apple Watch is an app that detects cardiac arrhythmias and even produces an electrocardiogram if an abnormality is detected. The technology, approved by the FDA, is helping cardiologists monitor heart patients and design interventions for those who may be at higher risk of a cardiac event like a stroke.
If having an algorithm predict your death sends a shiver down your spine, consider that algorithms may keep you alive longer. In 2018, technology reporter Tristan Greene wrote for Medium that "…despite the unending deluge of panic-ridden articles declaring AI the path to apocalypse, we're now living in a world where algorithms save lives every day. The future of machine learning isn't sentient killer robots. It's longer human lives."
The Risks of AI Compiling Your Data
To be sure, the advent of AI-infused medical technology is not without its risks. One risk is that the use of AI wearables constantly monitoring our vital signs could turn us into a nation of hypochondriacs, racing to our doctors every time there's a blip in some vital sign. Such a development could stress an already overburdened system that suffers from, among other things, a shortage of doctors and nurses. Another risk has to do with the privacy protections on the massive repository of intimately personal information that AI will have on us.
In an article recently published in the Journal of the American Medical Association, Australian researcher Kit Huckvale and colleagues examined the handling of data by 36 smartphone apps that assisted people with either depression or smoking cessation, two areas that could lend themselves to stigmatization if they fell into the wrong hands.
Out of the 36 apps, 33 shared their data with third parties, despite the fact that just 25 of those apps had a privacy policy at all and out of those, only 23 stated that data would be shared with third parties. The recipients of all that data? It went almost exclusively to Facebook and Google, to be used for advertising and marketing purposes. But there's nothing to stop it from ending up in the hands of insurers, background databases, or any other entity.
Even when data isn't voluntarily shared, any digital information can be hacked. EHRs and even wearable devices share the same vulnerability as any other digital record or device. Still, the promise of AI to radically improve efficiency and accuracy in healthcare is hard to ignore.
AI Can Help Restore Humanity to Medicine
Eric Topol, director of the Scripps Research Translational Institute and author of the new book Deep Medicine, says that AI gives doctors and nurses the most precious gift of all: time.
Topol welcomes his patients' use of the Apple Watch cardiac feature and is optimistic about the ways that AI is revolutionizing medicine. He says that the watch helps doctors monitor how well medications are working and has already helped to prevent strokes. But in addition to that, AI will help bring the humanity back to a profession that has become as cold and hard as a stainless steel dissection table.
"When I graduated from medical school in the 1970s," he says, "you had a really intimate relationship with your doctor." Over the decades, he has seen that relationship steadily erode as medical organizations demanded that doctors see more and more patients within ever-shrinking time windows.
"Doctors have no time to think, to communicate. We need to restore the mission in medicine."
In addition to that, EHRs have meant that doctors and nurses are getting buried in paperwork and administrative tasks. This is no doubt one reason why a recent study by the World Health Organization showed that worldwide, about 50 percent of doctors suffer from burnout. People who are utterly exhausted make more mistakes, and medical clinicians are no different from the rest of us. Only medical mistakes have unacceptably high stakes. According to its website, Johns Hopkins University recently announced that in the U.S. alone, 250,000 people die from medical mistakes each year.
"Doctors have no time to think, to communicate," says Topol. "We need to restore the mission in medicine." AI is giving doctors more time to devote to the thing that attracted them to medicine in the first place—connecting deeply with patients.
There is a real danger at this juncture, though, that administrators aware of the time-saving aspects of AI will simply push doctors to see more patients, read more tests, and embrace an even more crushing workload.
"We can't leave it to the administrators to just make things worse," says Topol. "Now is the time for doctors to advocate for a restoration of the human touch. We need to stand up for patients and for the patient-doctor relationship."
AI could indeed be a game changer, he says, but rather than squander the huge benefits of more time, "We need a new equation going forward."
Can Biotechnology Take the Allergies Out of Cats?
From a special food to a vaccine and gene editing, new technologies may offer solutions for cat lovers with allergies.
Amy Bitterman, who teaches at Rutgers Law School in Newark, gets enormous pleasure from her three mixed-breed rescue cats, Spike, Dee, and Lucy. To manage her chronically stuffy nose, three times a week she takes Allegra D, which combines the antihistamine fexofenadine with the decongestant pseudoephedrine. Amy's dog allergy is rougher--so severe that when her sister launched a business, Pet Care By Susan, from their home in Edison, New Jersey, they knew Susan would have to move elsewhere before she could board dogs. Amy has tried to visit their brother, who owns a Labrador Retriever, taking Allegra D beforehand. But she began sneezing, and then developed watery eyes and phlegm in her chest.
"It gets harder and harder to breathe," she says.
Animal lovers have long dreamed of "hypo-allergenic" cats and dogs. Although to date, there is no such thing, biotechnology is beginning to provide solutions for cat-lovers. Cats are a simpler challenge than dogs. Dog allergies involve as many as seven proteins. But up to 95 percent of people who have cat allergies--estimated at 10 to 30 percent of the population in North America and Europe--react to one protein, Fel d1. Interestingly, cats don't seem to need Fel d1. There are cats who don't produce much Fel d1 and have no known health problems.
The current technologies fight Fel d1 in ingenious ways. Nestle Purina reached the market first with a cat food, Pro Plan LiveClear, launched in the U.S. a year and a half ago. It contains Fel d1 antibodies from eggs that in effect neutralize the protein. HypoCat, a vaccine for cats, induces them to create neutralizing antibodies to their own Fel d1. It may be available in the United States by 2024, says Gary Jennings, chief executive officer of Saiba Animal Health, a University of Zurich spin-off. Another approach, using the gene-editing tool CRISPR to create a medication that would splice out Fel d1 genes in particular tissues, is the furthest from fruition.
"Our goal was to ensure that whatever we do has no negative impact on the cat."
Customer demand is high. "We already have a steady stream of allergic cat owners contacting us desperate to have access to the vaccine or participate in the testing program," Jennings said. "There is a major unmet medical need."
More than a third of Americans own a cat (while half own a dog), and pet ownership is rising. With more Americans living alone, pets may be just the right amount of company. But the number of Americans with asthma increases every year. Of that group, some 20 to 30 percent have pet allergies that could trigger a possibly deadly attack. It is not clear how many pets end up in shelters because their owners could no longer manage allergies. Instead, allergists commonly report that their patients won't give up a beloved companion.
No one can completely avoid Fel d1, which clings to clothing and lands everywhere cat-owners go, even in schools and new homes never occupied by cats. Myths among cat-lovers may lead them to underestimate their own level of risk. Short hair doesn't help: the length of cat hair doesn't affect the production of Fel d1. Bathing your cat will likely upset it and accomplish little. Washing cuts the amount on its skin and fur only for two days. In one study, researchers measured the Fel d1 in the ambient air in a small chamber occupied by a cat—and then washed the cat. Three hours later, with the cat in the chamber again, the measurable Fel d1 in the air was lower. But this benefit was gone after 24 hours.
For years, the best option has been shots for people that prompt protective antibodies. Bitterman received dog and cat allergy injections twice a week as a child. However, these treatments require up to 100 injections over three to five years, and, as in her case, the effect may be partial or wear off. Even if you do opt for shots, treating the cat also makes sense, since you could protect more than one allergic member of your household and any allergic visitors as well.
An Allergy-Neutralizing Diet
Cats produce much of their Fel d1 in their saliva, which then spreads it to their fur when they groom, observed Nestle Purina immunologist Ebenezer Satyaraj. He realized that this made saliva—and therefore a cat's mouth--an unusually effective site for change. Hens exposed to Fel d1 produce their own antibodies, which survive in their eggs. The team coated LiveClear food with a powder form of these eggs; once in a cat's mouth, the chicken antibody binds to the Fel d1 in the cat's saliva, neutralizing it.
The results are partial: In a study with 105 cats, the level of active Fel d1 in their fur had dropped on average by 47 percent after ten weeks eating LiveClear. Cats that produced more Fel d1 at baseline had a more robust response, with a drop of up to 71 percent. A safety study found no effects on cats after six months on the diet. "Our goal was to ensure that whatever we do has no negative impact on the cat," Satyaraj said. Might a dogfood that minimizes dog allergens be on the way? "There is some early work," he said.
A Vaccine
This is a year when vaccines changed the lives of billions. Saiba's vaccine, HypoCat, delivers recombinant Fel d1 and the coat from a plant virus (the Cucumber mosaic virus) without any vital genetic information. The viral coat serves as a carrier. A cat would need shots once or twice a year to produce antibodies that neutralize Fel d1.
HypoCat works much like any vaccine, with the twist that the enemy is the cat's own protein. Is that safe? Saiba's team has followed 70 cats treated with the vaccine over two years and they remain healthy. Again the active Fel d1 doesn't disappear but diminishes. The team asked 10 people with cat allergies to report on their symptoms when they pet their vaccinated cats. Eight of them could pet their cat for nearly a half hour before their symptoms began, compared with an average of 17 minutes before the vaccine.
Jennings hopes to develop a HypoDog shot with a similar approach. However, the goal would be to target four or five proteins in one vaccine, and that increases the risk of hurting the dog. In the meantime, allergic dog-lovers considering an expensive breeder dog might think again: Independent research does not support the idea that any breed of dog produces less dander in the home. In fact, one well-designed study found that Spanish water dogs, Airedales, poodles and Labradoodles--breeds touted as hypo-allergenic--had significantly more of the most common allergen on their coat than an ordinary Lab and the control group.
Gene Editing
One day you might be able to bring your cat to the vet once a year for an injection that would modify specific tissues so they wouldn't produce Fel d1.
Nicole Brackett, a postdoctoral scientist at Viriginia-based Indoor Biotechnologies, which specializes in manufacturing biologics for allergy and asthma, most recently has used CRISPR to identify Fel d1 genetic sequences in cells from 50 domestic cats and 24 exotic ones. She learned that the sequences vary substantially from one cat to the next. This discovery, she says, backs up the observations that Fel d1 doesn't have a vital purpose.
The next step will be a CRISPR knockout of the relevant genes in cells from feline salivary glands, a prime source of Fel d1. Although the company is considering using CRISPR to edit the genes in a cat embryo and possibly produce a Fel d1-free cat, designer cats won't be its ultimate product. Instead, the company aims to produce injections that could treat any cat.
Reducing pet allergens at home could have a compound benefit, Indoor Biotechnologies founder Martin Chapman, an immunologist, notes: "When you dampen down the response to one allergen, you could also dampen it down to multiple allergens." As allergies become more common around the world, that's especially good news.
The large genetic studies underlying certain disease risk tests have primarily been done in populations of European ancestry, limiting their accuracy.
Earlier this year, California-based Ambry Genetics announced that it was discontinuing a test meant to estimate a person's risk of developing prostate or breast cancer. The test looks for variations in a person's DNA that are known to be associated with these cancers.
Known as a polygenic risk score, this type of test adds up the effects of variants in many genes — often in the dozens or hundreds — and calculates a person's risk of developing a particular health condition compared to other people. In this way, polygenic risk scores are different from traditional genetic tests that look for mutations in single genes, such as BRCA1 and BRCA2, which raise the risk of breast cancer.
Traditional genetic tests look for mutations that are relatively rare in the general population but have a large impact on a person's disease risk, like BRCA1 and BRCA2. By contrast, polygenic risk scores scan for more common genetic variants that, on their own, have a small effect on risk. Added together, however, they can raise a person's risk for developing disease.
These scores could become a part of routine healthcare in the next few years. Researchers are developing polygenic risk scores for cancer, heart, disease, diabetes and even depression. Before they can be rolled out widely, they'll have to overcome a key limitation: racial bias.
"The issue with these polygenic risk scores is that the scientific studies which they're based on have primarily been done in individuals of European ancestry," says Sara Riordan, president of the National Society of Genetics Counselors. These scores are calculated by comparing the genetic data of people with and without a particular disease. To make these scores accurate, researchers need genetic data from tens or hundreds of thousands of people.
Myriad's old test would have shown that a Black woman had twice as high of a risk for breast cancer compared to the average woman even if she was at low or average risk.
A 2018 analysis found that 78% of participants included in such large genetic studies, known as genome-wide association studies, were of European descent. That's a problem, because certain disease-associated genetic variants don't appear equally across different racial and ethnic groups. For example, a particular variant in the TTR gene, known as V1221, occurs more frequently in people of African descent. In recent years, the variant has been found in 3 to 4 percent of individuals of African ancestry in the United States. Mutations in this gene can cause protein to build up in the heart, leading to a higher risk of heart failure. A polygenic risk score for heart disease based on genetic data from mostly white people likely wouldn't give accurate risk information to African Americans.
Accuracy in genetic testing matters because such polygenic risk scores could help patients and their doctors make better decisions about their healthcare.
For instance, if a polygenic risk score determines that a woman is at higher-than-average risk of breast cancer, her doctor might recommend more frequent mammograms — X-rays that take a picture of the breast. Or, if a risk score reveals that a patient is more predisposed to heart attack, a doctor might prescribe preventive statins, a type of cholesterol-lowering drug.
"Let's be clear, these are not diagnostic tools," says Alicia Martin, a population and statistical geneticist at the Broad Institute of MIT and Harvard. "We can't use a polygenic score to say you will or will not get breast cancer or have a heart attack."
But combining a patient's polygenic risk score with other factors that affect disease risk — like age, weight, medication use or smoking status — may provide a better sense of how likely they are to develop a specific health condition than considering any one risk factor one its own. The accuracy of polygenic risk scores becomes even more important when considering that these scores may be used to guide medication prescription or help patients make decisions about preventive surgery, such as a mastectomy.
In a study published in September, researchers used results from large genetics studies of people with European ancestry and data from the UK Biobank to calculate polygenic risk scores for breast and prostate cancer for people with African, East Asian, European and South Asian ancestry. They found that they could identify individuals at higher risk of breast and prostate cancer when they scaled the risk scores within each group, but the authors say this is only a temporary solution. Recruiting more diverse participants for genetics studies will lead to better cancer detection and prevent, they conclude.
Recent efforts to do just that are expected to make these scores more accurate in the future. Until then, some genetics companies are struggling to overcome the European bias in their tests.
Acknowledging the limitations of its polygenic risk score, Ambry Genetics said in April that it would stop offering the test until it could be recalibrated. The company launched the test, known as AmbryScore, in 2018.
"After careful consideration, we have decided to discontinue AmbryScore to help reduce disparities in access to genetic testing and to stay aligned with current guidelines," the company said in an email to customers. "Due to limited data across ethnic populations, most polygenic risk scores, including AmbryScore, have not been validated for use in patients of diverse backgrounds." (The company did not make a spokesperson available for an interview for this story.)
In September 2020, the National Comprehensive Cancer Network updated its guidelines to advise against the use of polygenic risk scores in routine patient care because of "significant limitations in interpretation." The nonprofit, which represents 31 major cancer cancers across the United States, said such scores could continue to be used experimentally in clinical trials, however.
Holly Pederson, director of Medical Breast Services at the Cleveland Clinic, says the realization that polygenic risk scores may not be accurate for all races and ethnicities is relatively recent. Pederson worked with Salt Lake City-based Myriad Genetics, a leading provider of genetic tests, to improve the accuracy of its polygenic risk score for breast cancer.
The company announced in August that it had recalibrated the test, called RiskScore, for women of all ancestries. Previously, Myriad did not offer its polygenic risk score to women who self-reported any ancestry other than sole European or Ashkenazi ancestry.
"Black women, while they have a similar rate of breast cancer to white women, if not lower, had twice as high of a polygenic risk score because the development and validation of the model was done in white populations," Pederson said of the old test. In other words, Myriad's old test would have shown that a Black woman had twice as high of a risk for breast cancer compared to the average woman even if she was at low or average risk.
To develop and validate the new score, Pederson and other researchers assessed data from more than 275,000 women, including more than 31,000 African American women and nearly 50,000 women of East Asian descent. They looked at 56 different genetic variants associated with ancestry and 93 associated with breast cancer. Interestingly, they found that at least 95% of the breast cancer variants were similar amongst the different ancestries.
The company says the resulting test is now more accurate for all women across the board, but Pederson cautions that it's still slightly less accurate for Black women.
"It's not only the lack of data from Black women that leads to inaccuracies and a lack of validation in these types of risk models, it's also the pure genomic diversity of Africa," she says, noting that Africa is the most genetically diverse continent on the planet. "We just need more data, not only in American Black women but in African women to really further characterize that continent."
Martin says it's problematic that such scores are most accurate for white people because they could further exacerbate health disparities in traditionally underserved groups, such as Black Americans. "If we were to set up really representative massive genetic studies, we would do a much better job at predicting genetic risk for everybody," she says.
Earlier this year, the National Institutes of Health awarded $38 million to researchers to improve the accuracy of polygenic risk scores in diverse populations. Researchers will create new genome datasets and pool information from existing ones in an effort to diversify the data that polygenic scores rely on. They plan to make these datasets available to other scientists to use.
"By having adequate representation, we can ensure that the results of a genetic test are widely applicable," Riordan says.