Monday, July 09, 2018

Is Dr. Google Ready to See Us Now?


Google’s parent company Alphabet has recently demonstrated an artificial-intelligence (AI) algorithm that can be used to estimate how likely hospital patients are to die soon.  A recent piece in Bloomberg News described how one particular woman with end-stage breast cancer arrived with fluid-filled lungs and underwent numerous tests.  The Google algorithm said there was a 20% chance she would not survive her hospital stay, and she died a few days later.

One data point—or one life.  The woman was both, and therein lies the challenge for researchers wanting to use AI to improve health care.  AI is a data-hungry beast, thriving on huge databases and sweeping up any scrap of information in its maw.  One of the best features of Google’s medical AI system is that it doesn’t need to have the raw data gussied up, in terms of needing a human being to type messy notes into a form the computer can use, a process that consumes as much as 80% of the effort devoted to other AI medical software.  Google’s system takes almost any kind of hand-scrawled data and integrates it into patient evaluations.  So in order to help, the system needs to know everything, no matter how apparently trivial or unrelated to the case it may be.

But then the human aspect enters.  To make my point, I’ll draw an analogy to a differetn profession—banking.  I’m old enough to remember when bankers evaluated customers with a combination of hard data—loans paid off in the past, bank balances, and so on—and intuition gained from meeting and talking with the customer.  Except for maybe a few high-class boutique banks, this is no longer the case.  The almighty credit score ground out by opaque algorithms reigns, and no amount of personal charm exerted for the benefit of a loan officer will overcome a low credit score. 

It’s one thing when we’re talking about loans, and another when the subject is human lives.  It’s easy to imagine a dystopian narrative involving a Google-like AI program that comes to dominate the decision-making process in a situation where medical resources are limited and there are more patients needing expensive care than the system can handle.  Doctors will turn to their AI assistants and ask, “Which of these five patients is most likely to benefit from a kidney transplant?”  It’s likely that some form of this process already goes on today, but is limited to comparatively rare situations such as transplants. 

The U. S. government’s Medicare system is currently forecast to become insolvent eight years from now.  Even if Congress manages to bail it out, the flood of aging baby-boomers such as myself will threaten to overwhelm the nation’s health-care system.  In such a crisis, the temptation to use AI algorithms to allocate limited resources will be overwhelming. 

From an engineering-efficiency standpoint, it all makes sense.  Why waste limited resources on someone who isn’t likely to benefit from them, when another person may get them and go on to live many years longer?  That’s fine except for two things.

One, even the best AI systems aren’t perfect, and now and then there will be mistakes—sometimes major ones.

And two, what if an AI medical system tells you you’re not going to get that treatment that might make the difference between life or death?  Even the hardiest utilitarian (“greatest benefit for the greatest number”) may have second thoughts about that outcome.

 Of course, resource allocation in health care is nothing new.  There have always been more sick people than there have been facilities to take care of them all.  The way we’ve done it in the past has been a combination of economics, the judgment of medical personnel, and government intervention from time to time.  As computers made inroads into various parts of the process, it’s only natural that they be used along with other available means to make wise choices.  But there’s a difference between using computers as tools and completely turning over decision-making to an algorithm.

Another concern raised about Google’s foray into applying AI to health care is the issue of privacy.  Medical records are among the most sensitive types of personal data, and in the past, elaborate precautions have been taken to guard the sanctity of each individual’s records.  But AI algorithms work better the more data they have, and so simply for the purpose of getting better at what they do, these algorithms will need access to as much data as they can get their digital hands on.  According to one survey, less than half of the public trusts Google to keep their data private.  While that is just a perception, it’s a perception that Google, and the medical profession in general, ignore at their peril.  One scandal or major data breach involving medical records could set back the entire medical-AI industry, so all participants will need to tread carefully and make sure nothing like that happens, or else the whole experiment could come to a screeching halt.

Predicting when people will die is only one of the many abilities that medical AI of the future offers.  In figuring out hard-to-diagnose cases, in recommending treatment customized to the individual, and in optimizing the delivery of health care generally, it shows great promise in making health care more effective and efficient for the vast majority of patients.  But doctors and other medical authorities should beware of letting the algorithms gain the upper hand, and turning their judgment and ethics over to a printout, so to speak.  Because Google’s system is still in the prototype stage, we don’t know what the effects of its more widespread deployment will be.  But whatever form it takes, we need to make sure that the vital life-or-death decisions involved in medical care are made by responsible people, not just machines.

Sources:  The article “Google Is Training Machines to Predict When a Patient Will Die” appeared on June 18, 2018 in Bloomberg News at https://www.bloomberg.com/news/articles/2018-06-18/google-is-training-machines-to-predict-when-a-patient-will-die, and was reprinted by the Austin American-Statesman, where I first saw it.  I also referred to an article on the Modern Health Care website at http://www.modernhealthcare.com/article/20180419/NEWS/180419911.  The statistic about Medicare’s insolvency is from p. 6 of the July 9, 2018 issue of National Review.   

No comments:

Post a Comment