Whenever Eric Karl Oermann has to tell a patient about a terrible prognosis, their first question is always: "how long do I have?" Oermann would like to offer a precise answer, to provide some certainty and help guide treatment. But although he's one of the country's foremost experts in medical artificial intelligence, Oermann is still dependent on a computer algorithm that's often wrong.
Doctors are notoriously terrible at guessing how long their patients will live.
Artificial intelligence, now often called deep learning or neural networks, has radically transformed language and image processing. It's allowed computers to play chess better than the world's grand masters and outwit the best Jeopardy players. But it still can't precisely tell a doctor how long a patient has left – or how to help that person live longer.
Someday, researchers predict, computers will be able to watch a video of a patient to determine their health status. Doctors will no longer have to spend hours inputting data into medical records. And computers will do a better job than specialists at identifying tiny tumors, impending crises, and, yes, figuring out how long the patient has to live. Oermann, a neurosurgeon at Mount Sinai, says all that technology will allow doctors to spend more time doing what they do best: talking with their patients. "I want to see more deep learning and computers in a clinical setting," he says, "so there can be more human interaction." But those days are still at least three to five years off, Oermann and other researchers say.
Doctors are notoriously terrible at guessing how long their patients will live, says Nigam Shah, an associate professor at Stanford University and assistant director of the school's Center for Biomedical Informatics Research. Doctors don't want to believe that their patient – whom they've come to like – will die. "Doctors over-estimate survival many-fold," Shah says. "How do you go into work, in say, oncology, and not be delusionally optimistic? You have to be."
But patients near the end of life will get better treatment – and even live longer – if they are overseen by hospice or palliative care, research shows. So, instead of relying on human bias to select those whose lives are nearing their end, Shah and his colleagues showed that they could use a deep learning algorithm based on medical records to flag incoming patients with a life expectancy of three months to a year. They use that data to indicate who might need palliative care. Then, the palliative care team can reach out to treating physicians proactively, instead of relying on their referrals or taking the time to read extensive medical charts.
But, although the system works well, Shah isn't yet sure if such indicators actually get the appropriate patients into palliative care. He's recently partnered with a palliative care doctor to run a gold-standard clinical trial to test whether patients who are flagged by this algorithm are indeed a better match for palliative care.
"What is effective from a health system perspective might not be effective from a treating physician's perspective and might not be effective from the patient's perspective," Shah notes. "I don't have a good way to guess everybody's reaction without actually studying it." Whether palliative care is appropriate, for instance, depends on more than just the patient's health status. "If the patient's not ready, the family's not ready and the doctor's not ready, then you're just banging your head against the wall," Shah says. "Given limited capacity, it's a waste of resources" to put that person in palliative care.
The algorithm isn't perfect, but "on balance, it leads to better decisions more often."
Alexander Smith and Sei Lee, both palliative care doctors, work together at the University of California, San Francisco, to develop predictions for patients who come to the hospital with a complicated prognosis or a history of decline. Their algorithm, they say, helps decide if this patient's problems – which might include diabetes, heart disease, a slow-growing cancer, and memory issues – make them eligible for hospice. The algorithm isn't perfect, they both agree, but "on balance, it leads to better decisions more often," Smith says.
Bethany Percha, an assistant professor at Mount Sinai, says that an algorithm may tell doctors that their patient is trending downward, but it doesn't do anything to change that trajectory. "Even if you can predict something, what can you do about it?" Algorithms may be able to offer treatment suggestions – but not what specific actions will alter a patient's future, says Percha, also the chief technology officer of Precise Health Enterprise, a product development group within Mount Sinai. And the algorithms remain challenging to develop. Electronic medical records may be great at her hospital, but if the patient dies at a different one, her system won't know. If she wants to be certain a patient has died, she has to merge social security records of death with her system's medical records – a time-consuming and cumbersome process.
An algorithm that learns from biased data will be biased, Shah says. Patients who are poor or African American historically have had worse health outcomes. If researchers train an algorithm on data that includes those biases, they get baked into the algorithms, which can then lead to a self-fulfilling prophesy. Smith and Lee say they've taken race out of their algorithms to avoid this bias.
Age is even trickier. There's no question that someone's risk of illness and death goes up with age. But an 85-year-old who breaks a hip running a marathon should probably be treated very differently than an 85-year-old who breaks a hip trying to get out of a chair in a dementia care unit. That's why the doctor can never be taken out of the equation, Shah says. Human judgment will always be required in medical care and an algorithm should never be followed blindly, he says.
Experts say that the flaws in artificial intelligence algorithms shouldn't prevent people from using them – carefully.
Researchers are also concerned that their algorithms will be used to ration care, or that insurance companies will use their data to justify a rate increase. If an algorithm predicts a patient is going to end up back in the hospital soon, "who's benefitting from knowing a patient is going to be readmitted? Probably the insurance company," Percha says.
Still, Percha and others say, the flaws in artificial intelligence algorithms shouldn't prevent people from using them – carefully. "These are new and exciting tools that have a lot of potential uses. We need to be conscious about how to use them going forward, but it doesn't mean we shouldn't go down this road," she says. "I think the potential benefits outweigh the risks, especially because we've barely scratched the surface of what big data can do right now."
The Friday Five covers five stories in health research that you may have missed this week. There are plenty of controversies and troubling ethical issues in science – and we get into many of them in our online magazine – but this news roundup focuses on scientific creativity and progress to give you a therapeutic dose of inspiration headed into the weekend.
One day in recent past, scientists at Columbia University’s Creative Machines Lab set up a robotic arm inside a circle of five streaming video cameras and let the robot watch itself move, turn and twist. For about three hours the robot did exactly that—it looked at itself this way and that, like toddlers exploring themselves in a room full of mirrors. By the time the robot stopped, its internal neural network finished learning the relationship between the robot’s motor actions and the volume it occupied in its environment. In other words, the robot built a spatial self-awareness, just like humans do. “We trained its deep neural network to understand how it moved in space,” says Boyuan Chen, one of the scientists who worked on it.
For decades robots have been doing helpful tasks that are too hard, too dangerous, or physically impossible for humans to carry out themselves. Robots are ultimately superior to humans in complex calculations, following rules to a tee and repeating the same steps perfectly. But even the biggest successes for human-robot collaborations—those in manufacturing and automotive industries—still require separating the two for safety reasons. Hardwired for a limited set of tasks, industrial robots don't have the intelligence to know where their robo-parts are in space, how fast they’re moving and when they can endanger a human.
Over the past decade or so, humans have begun to expect more from robots. Engineers have been building smarter versions that can avoid obstacles, follow voice commands, respond to human speech and make simple decisions. Some of them proved invaluable in many natural and man-made disasters like earthquakes, forest fires, nuclear accidents and chemical spills. These disaster recovery robots helped clean up dangerous chemicals, looked for survivors in crumbled buildings, and ventured into radioactive areas to assess damage.
Now roboticists are going a step further, training their creations to do even better: understand their own image in space and interact with humans like humans do. Today, there are already robot-teachers like KeeKo, robot-pets like Moffin, robot-babysitters like iPal, and robotic companions for the elderly like Pepper.
But even these reasonably intelligent creations still have huge limitations, some scientists think. “There are niche applications for the current generations of robots,” says professor Anthony Zador at Cold Spring Harbor Laboratory—but they are not “generalists” who can do varied tasks all on their own, as they mostly lack the abilities to improvise, make decisions based on a multitude of facts or emotions, and adjust to rapidly changing circumstances. “We don’t have general purpose robots that can interact with the world. We’re ages away from that.”
Robotic spatial self-awareness – the achievement by the team at Columbia – is an important step toward creating more intelligent machines. Hod Lipson, professor of mechanical engineering who runs the Columbia lab, says that future robots will need this ability to assist humans better. Knowing how you look and where in space your parts are, decreases the need for human oversight. It also helps the robot to detect and compensate for damage and keep up with its own wear-and-tear. And it allows robots to realize when something is wrong with them or their parts. “We want our robots to learn and continue to grow their minds and bodies on their own,” Chen says. That’s what Zador wants too—and on a much grander level. “I want a robot who can drive my car, take my dog for a walk and have a conversation with me.”
Columbia scientists have trained a robot to become aware of its own "body," so it can map the right path to touch a ball without running into an obstacle, in this case a square.
Jane Nisselson and Yinuo Qin/ Columbia Engineering
Today’s technological advances are making some of these leaps of progress possible. One of them is the so-called Deep Learning—a method that trains artificial intelligence systems to learn and use information similar to how humans do it. Described as a machine learning method based on neural network architectures with multiple layers of processing units, Deep Learning has been used to successfully teach machines to recognize images, understand speech and even write text.
Trained by Google, one of these language machine learning geniuses, BERT, can finish sentences. Another one called GPT3, designed by San Francisco-based company OpenAI, can write little stories. Yet, both of them still make funny mistakes in their linguistic exercises that even a child wouldn’t. According to a paper published by Stanford’s Center for Research on Foundational Models, BERT seems to not understand the word “not.” When asked to fill in the word after “A robin is a __” it correctly answers “bird.” But try inserting the word “not” into that sentence (“A robin is not a __”) and BERT still completes it the same way. Similarly, in one of its stories, GPT3 wrote that if you mix a spoonful of grape juice into your cranberry juice and drink the concoction, you die. It seems that robots, and artificial intelligence systems in general, are still missing some rudimentary facts of life that humans and animals grasp naturally and effortlessly.
How does one give robots a genome? Zador has an idea. We can’t really equip machines with real biological nucleotide-based genes, but we can mimic the neuronal blueprint those genes create.
It's not exactly the robots’ fault. Compared to humans, and all other organisms that have been around for thousands or millions of years, robots are very new. They are missing out on eons of evolutionary data-building. Animals and humans are born with the ability to do certain things because they are pre-wired in them. Flies know how to fly, fish knows how to swim, cats know how to meow, and babies know how to cry. Yet, flies don’t really learn to fly, fish doesn’t learn to swim, cats don’t learn to meow, and babies don’t learn to cry—they are born able to execute such behaviors because they’re preprogrammed to do so. All that happens thanks to the millions of years of evolutions wired into their respective genomes, which give rise to the brain’s neural networks responsible for these behaviors. Robots are the newbies, missing out on that trove of information, Zador argues.
A neuroscience professor who studies how brain circuitry generates various behaviors, Zador has a different approach to developing the robotic mind. Until their creators figure out a way to imbue the bots with that information, robots will remain quite limited in their abilities. Each model will only be able to do certain things it was programmed to do, but it will never go above and beyond its original code. So Zador argues that we have to start giving robots a genome.
How does one do that? Zador has an idea. We can’t really equip machines with real biological nucleotide-based genes, but we can mimic the neuronal blueprint those genes create. Genomes lay out rules for brain development. Specifically, the genome encodes blueprints for wiring up our nervous system—the details of which neurons are connected, the strength of those connections and other specs that will later hold the information learned throughout life. “Our genomes serve as blueprints for building our nervous system and these blueprints give rise to a human brain, which contains about 100 billion neurons,” Zador says.
If you think what a genome is, he explains, it is essentially a very compact and compressed form of information storage. Conceptually, genomes are similar to CliffsNotes and other study guides. When students read these short summaries, they know about what happened in a book, without actually reading that book. And that’s how we should be designing the next generation of robots if we ever want them to act like humans, Zador says. “We should give them a set of behavioral CliffsNotes, which they can then unwrap into brain-like structures.” Robots that have such brain-like structures will acquire a set of basic rules to generate basic behaviors and use them to learn more complex ones.
Currently Zador is in the process of developing algorithms that function like simple rules that generate such behaviors. “My algorithms would write these CliffsNotes, outlining how to solve a particular problem,” he explains. “And then, the neural networks will use these CliffsNotes to figure out which ones are useful and use them in their behaviors.” That’s how all living beings operate. They use the pre-programmed info from their genetics to adapt to their changing environments and learn what’s necessary to survive and thrive in these settings.
For example, a robot’s neural network could draw from CliffsNotes with “genetic” instructions for how to be aware of its own body or learn to adjust its movements. And other, different sets of CliffsNotes may imbue it with the basics of physical safety or the fundamentals of speech.
At the moment, Zador is working on algorithms that are trying to mimic neuronal blueprints for very simple organisms—such as earthworms, which have only 302 neurons and about 7000 synapses compared to the millions we have. That’s how evolution worked, too—expanding the brains from simple creatures to more complex to the Homo Sapiens. But if it took millions of years to arrive at modern humans, how long would it take scientists to forge a robot with human intelligence? That’s a billion-dollar question. Yet, Zador is optimistic. “My hypotheses is that if you can build simple organisms that can interact with the world, then the higher level functions will not be nearly as challenging as they currently are.”