Google Testing AI That Can Predict When You Die
Using a proto-AI neural network, Google is experimenting with ways for programs to predict when you die, how long it will take, and more.
Almost since the first experiments with machine learning and neural networks, companies have been looking at two things: how to make the most of these tools and how to monetize them. Google thinks it may have found an answer that satisfies both.
In a recent study that focused on a woman diagnosed with late-stage breast cancer, two computer models assessed the woman’s chances of survival. The hospital’s own computer systems factored in a radiology scan, her vital signs, and that she already had fluid in her lungs. It gave her a 9.3 percent chance that she would leave the hospital alive. The Google model, which considered 175,693 data points, gave her a slightly higher chance at 19.9 percent, but only a slightly larger chance.
The patient died within 10 days of the predictions. And while it’s slightly ghoulish to celebrate that fact, it did hint at a huge amount of potential for neural networks, and eventually true AI in the field of healthcare.
Using experimental technology like machine learning in healthcare isn’t a new idea, but where Google is beginning to pull away from others is in how its neural net can consume data. When it comes to predictive models, nearly 80 percent of the time spent working on them comes down to “scut work,” which means making the data presentable. It is also costly and requires finesse. With Google’s program, you can “throw in the kitchen sink and not have to worry about it,” according to Nigam Shah, an associate professor at Stanford and co-author of Google’s latest research paper on the subject.
Along with swallowing up huge amounts of data and sorting it out, the predictive model can also highlight where the information came from and flag certain things. In theory, this will lead to an AI one day being able to inform medical staff about things they may have missed, and eventually predict the outcome of a patient. That would make it an indispensable – and extremely profitable – tool for hospital and healthcare providers. And even if the predictive portion of the program never works, the data collection ability alone would save medical facilities time and money. That makes it potentially valuable as is, but with a little more work it will be able to go much further.
Before it reaches the point where it can look at a patient and make a diagnosis, Google first needs to figure out how to make different forms of data universal and sharable. As part of its testing, Google provided software to two hospitals. Both were able to make use of the tool in terms of records, but the different idiosyncrasies from each hospital’s systems made it almost impossible to share data between the two. Creating something that would work across all hospital and accept data from all medical systems, even those created and coded by hand, will be a daunting task – but not an insurmountable one.
Once that happens and as the neural network becomes a true AI and can begin to make its own connections, the medical field could change rapidly. Add in data from personal tracking devices and things get even more interesting. Imagine your Fitbit monitoring and storing all your vitals and then uploading them to a medical sensor, and the AI might be able to see things that your doctor would never detect. And if you were brought into a hospital unresponsive, the AI could instantly review your medical history and your bio information over the last few months or even years, and have a basic treatment plan ready to go before a doctor even saw you.
There are, of course, several potential downsides to this.
To start with, Google has already run afoul of privacy groups several times before, and mixing your personal data with your biodata might be a little much for some of us. It’s one thing to see a targetted ad for shoes, it’s another to receive that ad with a warning that if you don’t start jogging to counteract the burger you just ate you could die within the next five years based. Last year, another Google neural network, DeepMind, was punished by British regulators for testing an app that went through public medical records without telling the people what they were doing.
Then there are insurance companies. If the AI is good enough at diagnosing patients and sets a treatment schedule complete with a timetable, if the patient’s recovery isn’t fast enough, or if they don’t die “on schedule,” the insurance companies might find a way to penalize the patient financially. That might be a little alarmist, but if you’ve had extensive dealings with an insurance company lately, it probably won’t seem that far-fetched.
For now, however, the neural nets are likely to be relegated to data collection and sorting. Soon though, they will be given larger tasks. And from there, the future is AI.