Death is an uncomfortable subject for many people, and yet it the question of our own mortality seems to be a common search query. Just enter the phrase “when will” into Google, and “when will I die” is currently the third search suggestion, just behind “when will the world be free” and “when will iOS 12 be released.” But can Google seriously tell you when you are going to die?
Personally, I find the idea that an Internet search engine can predict an individual’s death with any sort of accuracy to be ludicrous. There are just too many variables. I mean sure, a search engine might theoretically be able to use statistical data to narrow things down to a 15 or 20 year window, or it might be able to “predict” someone’s demise based on a publicly announced execution date, but that’s about the extent of what a search engine might be reasonably expected to predict.
Out of curiosity, I did a Google search on the phrase “when will I die.” As expected, Google did not return a date. Instead, it listed a whole bunch of web pages, many of which claim to be able to accurately predict how long you will live based on how you complete a short quiz. So that’s the end of the story, right? Not quite.
The results above are Google’s search results based on the query “when will I die.”
A couple of weeks ago, I had to write a paper for a class that I am taking on spacecraft life support systems. While scouring the Internet for information, I stumbled onto an article titled “Google is training machines to predict when a patient will die.”
Morbid Google algorithms
The article cites a case in which an algorithm, which was created by Google, evaluated 175,639 unique data points and used that information to determine the odds that a particular cancer patient would survive their stay in the hospital. In essence, Google is trying to use machine learning to create a system that can accurately predict the outcomes of hospital patients.
Now I could spend the rest of this article rambling on about how the technology works, or how someone might use a similar technique (and a cloud-based computational cluster) to handicap horse races, but I think that there are some other aspects to this story that are far more interesting.
Let me just say upfront that I am skeptical of Google’s ability to accurately predict patient mortality. Statistical analytics can certainly be used to render probabilities, but there is a huge difference between probability and certainty. To put this into perspective, let’s imagine for a moment that a gravely ill patient is admitted to the hospital and that the world’s most advanced AI system gives that patient five nines of certainty (a 99.999 percent chance) of dying within the next 24 hours. In spite of having a top-notch algorithm and mountains of data available to it, there are still factors that the AI engine cannot account for. Some of these factors might include an especially skilled doctor who is able to beat the odds, divine intervention, or perhaps the unexpected availability of an experimental drug. The point is that software is great for calculating statistical probabilities, but it lacks the ability to truly see the future.
I think that any time that technology can be used to improve the quality of health care, it can only be a good thing. I am totally willing to give Google the benefit of a doubt and assume that their intentions are good. Even so, I can think of at least two ways in which Google’s efforts could backfire.
First, I can’t help but wonder if Google’s algorithm might eventually shape triage procedures. In medicine, triage is the practice of assessing patient’s needs based on the urgency of treatment. In an emergency room, for example, a patient who is experiencing chest pains will take priority over a patient who has a broken arm. The most critical patients are treated first — most of the time.
The exception to the rule
The exception to the rule occurs when the number of patients is far greater than what the medical staff can realistically handle. In those situations, a gravely ill patient might not receive medical care if it is determined that little can be done for them and that there are other patients who urgently need treatment. This practice is especially common in mass casualty events.
Given the way that triage works, I can’t help but wonder if some patients could be neglected because Google says that they are going to die anyway. As far as I know, this has not actually happened. I am just speculating.
Another thing that I can’t help but wonder about is spinoff technologies. If Google is analyzing vast amounts of patient data in an effort to determine the patient’s odds of survival, then there are other things that could presumably also be gleaned from the data. For example, an AI engine might determine that the patient will survive, but that there is a high probability that the patient will be readmitted three weeks from now. Such software might also be able to anticipate the eventual onset of a chronic condition based on lifestyle factors, past medical history, and current physical condition.
If software were eventually able to accurately make these types of projections, it is something that insurers would undoubtedly be very interested in. I can just imagine someone’s premiums going up because a computer projects that they will become diabetic in about seven years.
The thing that I find most interesting about the work that Google is doing is the data sources that are being used. As we all know, AI cannot work without good data, and ingesting larger quantities of credible data usually renders better results. But what types of data is the Google engine actually ingesting?
I am assuming that Google is leveraging electronic health records, but according to the article that I cited earlier, “What impressed medical experts most was Google’s ability to sift through data previously out of reach: notes buried in PDFs or scribbled on old charts. The neural net gobbled up all this unruly information then spat out predictions. And it did it far faster and more accurately than existing techniques. Google’s system even showed which records led it to conclusions.”
The fact that Google’s engine is able to not only parse but make sense of so many different types of disparate data is huge. It has the potential to be used to present physicians with a single source of the truth, and will almost certainly save both lives and patient frustration. Let me give you an example.
One of the more difficult diseases for physicians to accurately diagnose is lupus. Its symptoms are easily mistaken for other illnesses, and there are also certain types of viral infections that can partially mimic lupus.
The American College of Rheumatology has outlined eleven different symptoms for lupus, and the criteria that it has established for diagnosing a patient with lupus is that the patient has to have four of the eleven symptoms. But here is the problem. The four symptoms do not have to all occur simultaneously.
So with that in mind, imagine that a patient’s electronic health records include past references to two or three lupus symptoms. That alone is not sufficient for diagnosing lupus. The patient might, therefore, be misdiagnosed as having rheumatoid arthritis. But what if there was a handwritten note on an old medical chart (predating electronic health records) indicating that the patient has had additional Lupus symptoms? Today, such a note would almost certainly be overlooked, leading to a misdiagnosis. However, if Google is able to ingest and understand unstructured, possibly even handwritten data, then it could make it far easier for a physician to make an accurate diagnosis.
The good and the bad
Personally, I am not a big fan of Google’s quest to make predictions about whether a patient will live or die. Even so, I have long said that technology is neither good nor evil. It is how it is used that counts. While Google’s AI could very easily be abused or used to jack up our health insurance premiums, it also has the potential to save lives. Let’s just hope for all of our sakes that the technology is used responsibly.
Featured image: Shutterstock