IBM's Watson—the same machine that beat Ken Jennings at Jeopardy—is now churning through case histories at Memorial Sloan-Kettering, learning to make diagnoses and treatment recommendations. This is one in a series of developments suggesting that technology may be about to disrupt health care in the same way it has disrupted so many other industries. Are doctors necessary? Just how far might the automation of medicine go?
(The Atlantic March 2013).
The genesis of many of the efforts in medical informatics over the 30(+) years of development have centered on the grandiose efforts to model and then replace the practical intelligence of the physician with automated logic. Most recognized of these efforts is probably Mycin the rule-based expert system developed by Dr. Edward Shortliffe as his PhD project in the early 1970's. However, as revolutionary as the rule-based inference engine that was at the core of the development of mycin, and as widespread as that engine subsequently became as the reasoning core for expert system 'shells', most notably KEE (Knowledge Engineering Environment) it was never implemented in a clinically meaningful form.
Now enters Watson, the winner of the renowned Jeopardy match of the century... a truly impressive tool that is able to digest 60 million pages of text per second, being engineered to enhance the clinical decision making of mere humans, by assuring that they are aware of all the ground breaking 'research' within the field of their practice. In the Atlantic article cited above, the author breathlessly elaborates a clinical scenario in which the human oncologist did not test the patient for the KRAS variant of pulmonary adenocarcinoma and in fact Watson recognized this fact and made such a recommendation.
Let me quote in full from the article:
"Harley lukov didn’t need a miracle. He just needed the right diagnosis. Lukov, a 62-year-old from central New Jersey, had stopped smoking 10 years earlier—fulfilling a promise he’d made to his daughter, after she gave birth to his first grandchild. But decades of cigarettes had taken their toll. Lukov had adenocarcinoma, a common cancer of the lung, and it had spread to his liver. The oncologist ordered a biopsy, testing a surgically removed sample of the tumor to search for particular “driver” mutations. A driver mutation is a specific genetic defect that causes cells to reproduce uncontrollably, interfering with bodily functions and devouring organs. Think of an on/off switch stuck in the “on” direction. With lung cancer, doctors typically test for mutations called EGFR and ALK, in part because those two respond well to specially targeted treatments. But the tests are a long shot: although EGFR and ALK are the two driver mutations doctors typically see with lung cancer, even they are relatively uncommon. When Lukov’s cancer tested negative for both, the oncologist prepared to start a standard chemotherapy regimen—even though it meant the side effects would be worse and the prospects of success slimmer than might be expected using a targeted agent.
But Lukov’s true medical condition wasn’t quite so grim. The tumor did have a driver—a third mutation few oncologists test for in this type of case. It’s called KRAS. Researchers have known about KRAS for a long time, but only recently have they realized that it can be the driver mutation in metastatic lung cancer—and that, in those cases, it responds to the same drugs that turn it off in other tumors. A doctor familiar with both Lukov’s specific medical history and the very latest research might know to make the connection—to add one more biomarker test, for KRAS, and then to find a clinical trial testing the efficacy of KRAS treatments on lung cancer. But the national treatment guidelines for lung cancer don’t recommend such action, and few physicians, however conscientious, would think to do these things."
Now lets ignore several things here, i.e., that pulmonary adenocarcinoma is equally likely to affect non-smokers as smokers, i.e., it is not related to the 'risk factor' of tobacco ingestion, that national guidelines for treatment are national guidelines related to the 'best evidence' related to that illness at the time of their development, that the belief that enrollment in an oncological trial means a superior result than conventional therapy is naive, flawed, and demonstrates a gross misunderstanding of the process of clinical trial research. Given all these caveats I am less than impressed with the performance of Watson in this particular case. Specifically, in the case of advanced metastatic cancer of any cause, enrollment and personal outcome in a clinical trial is completely dependent on not only the 'right diagnosis' but also in the luck associated with contributing to the right clinical trial and being lucky enough to be in the right arm of that study.
What seems to be at play here, is we are asking the wrong question or using the tool in the wrong way. Given the power of the instrument, rather than attempting to unsuccessfully compete as a clinical expert, instead providing real time cost effective support for the housekeeping chores increasingly required of practitioners seems the more useful way to use this type of computing power. How cheaply could Watson function as a natural language translator of dictation and documentation or machine powered accurate diagnosis coding? Not sexy, not 'cutting edge' but in a health care system dominated by a 33% 'admins-trivia' carrying cost, i.e., 1/3 of all health care expenses are spent on the administration of the system rather than in direct patient care - how profound could the cost savings accrue if we could automate the non-clinical value added activities of coding and transcription and documentation of care?
In my ideal world, all clinical encounters would be videotaped and anyone interested in anything other than the summarization of the clinical encounter by the clinical expert could review the taped encounter for whatever picayune requirement that 3rd party might attempt to impose for their unique administrative requirements. Watson could make an immediate and real impact on clinical care, however, such transparent facilitation of the care of existing providers has a much lower priority than the 'pie in the sky' attempts to capture and emulate clinical expertise by machine driven approaches, an approach and goal that has littered and corrupted the entire field of medical informatics from its earliest days.