The Turing Test Falls: Implications for Health Care Decision Support

In the futuristic movie Blade Runner, Detective Rick Deckard's (played by Harrison Ford) skill at "retiring" renegade robotic replicants depends on a series of trick questions that are designed to detect an "empathic" response. While the soulless robots routinely fail the test, the highly advanced Nexus-6 models still seem to be eerily human. While Deckard violently terminates three of the robots, lingering questions over just what is "human" leads him to fall for vulnerable sexy replicant Rachael.

While the Population Health Blog ponders that, along comes the news that a Russian chatbot computer passed the Turing test. More than 30% of the humans who engaged in a text-only "conversation" with the program thought it was being controlled by a 13 year old boy. Not only was the computer able to organize facts and sentences, it also responded with the subtle nuances that underlie typical "human" communication.

While the PHB is weirded out, it is not surprised. In the book The Second Machine Age, authors Erik Brynjolfsson and Andrew McAfee note the doubling of computers' processing power can be likened to the ancient story of doubling wheat seeds on the squares of a chess board.  They point out that the amount of wheat (or processing power) can be grasped until you get to the "second half" of the board: that's when the amounts become staggering and the implications start getting weird.

They point out that computing power has now entered that second half. Quadruped "mule pack" machines can carry payloads across unfriendly landscapes, entire factories can manufacture complex items at a fraction of the cost and Watson can win Jeopardy matches. 

And now, Turing has fallen.

This is good news for health care.  "Second half" decision support in electronic health records is better able to focus on a more likely differential diagnosis, suggest a more accurate series of tests and tailor treatment at the point of care. The good news is that medicine will finally become faster, better and cheaper.  While some may fret about the loss of the "human touch" (or jobs) in this brave new world of the doctor-patient relationship, Brynjolfsson and McAfee point out that when human intelligence is combined with the resources of high performing information technology, the product is better than either alone.  For example, a chess master plus a high-end chess program can beat either alone. 

The same will be true in medicine: smart doctors plus nuanced health information technology will be better than either alone.

Just like in Blade Runner. Thanks to each other, both Deckard and Rachael are better... humans.

Image from Wikipedia

Electronic Health Record (EHR) "Robots" and The Need for a Legal Framework

Hal of 2001: A Space Odyssey

The Economist's recent scrutiny of battlefield and automotive robotics prompted the Disease Management Care Blog to ponder the parallels of "robot-like" health care. 

Sounds too weird, you say?  Alarmist?  Silly?  Read on.

Consider the following scenario:

After checking in with her primary care provider, a check of Mrs. Smith's (not her real name) diabetes revealed her blood sugar was not in the generally recommended A1c range of 6.5-7% (see page 16).  She was free of low blood sugar spells and was otherwise healthy. The electronic health record (EHR) decision support protocols (like "Diabetes Wizard") prompted the care manager to OK a standard increase in Mrs. Smith's insulin dose ("Share the Care"). This was not only consistent with good clinical practice, the decision promised a pay-for-performance payday for the doc's employer and fewer costly complications for the health insurer.
 
Several days later, her fatal car accident was ascribed to low blood sugar.  It turns out that the science of blood sugar control is still evolving, because tight diabetes control isn't necessary associated with better outcomes. In response, the Smith family lawyered-up.

The Economist article asks about the legal dilemmas inherent in robotic decision logic.  Can battlefield drones be adequately programmed to protect civilians?  Can driverless cars intentionally run over a dog if it means avoiding a child? Short of banning these automatons entirely, it's obvious, says The Economist, that we're going to have to develop a legal framework to deal with them: 

"...laws are needed to determine whether the designer, the programmer, the manufacturer or the operator is at fault if an autonomous drone strike goes wrong or a driverless car has an accident."

The DMCB agrees and suggests that this sensible recommendation be extended to the health care arena. EHR clinical programming may only imperfectly reconcile maximum near-term safety with longer term outcomes and fail to discern the subtleties that favor one treatment over another.

While white-coated plastic and metal cybertrons aren't taking care of diabetic Mrs. Smiths (yet), machine logic reminiscent of "Hal 9000" of 2001: A Space Odyssey already is. Instead of an unblinking red orb, there are on-screen decision-support icons. Instead of a disembodied voice, there are clinical pathways (for example, see page 8). And instead of having control over vacuum locks that exit to space, these computer entities are using logic trees to direct, prompt and recommend that doctors and their team-nurses to chose treatment "A" over treatment "B."  If you think that the DMCB is being alarmist, consider that decision support, at last count, is involved in 17% of all outpatient visits in the U.S.  That's a lot of legal exposure.

That's why reasonable laws are needed here too, since it's unclear how physicians, care managers, decision support programmers, EHR manufacturers (even with their notorious hold harmless clauses), physician clinics and health insurers would share in the culpability for allegedly awry decision-making.  Since, in the current U.S. tort system, all could be named in a suit, the risk is arguably intolerable. Just ask any physician, whose dread of being sued has already resulted in billions of unnecessary costs. 

Allowing that dysfunction to creep into computer-based clinical decision will increase costs there too, add uncertainty and slow adoption. 

The Economist has it right.

Image from Wikipedia