Watson Replacing Radiologists?

I like reading the Health Care Blog.  It typifies what is wrong with the management of the American Health Care system and I suppose blogs in general.  It is a steady stream of bad ideas and political rhetoric.  The best recent example was a little piece about radiologists called Will Watson Replace A Radiologist - Ask A Radiologist.  Radiologists either don't read this blog or they can't be bothered since the only comment at this point is from a rheumatologist on the necessary consultation and collegiality with radiologists.  The author of the main article is taking the perspective of being both threatening (Can the IBM Watson machine acquire the image reading capabilities of a human radiologist by "reading" a large set of clinical images and reading them at a much faster rate than a radiologist?) and advising (The only way that radiology will survive is to demonstrate their value to patients and colleagues by connecting with them?).  The author's conclusion is very explicit: Connect or be replaced.

Over the past thirty years my experience with radiologists has been positive and in some cases outstanding.  That dates back to the early days of being the medical student or intern responsible for carrying a stack of heavy and awkward films around.  I remember not having a film on a Cardiology rotation and regretting it: "Mr. Dawson - what made you think it was not a good idea to have the chest x-ray of this patient with mitral valve disease?"  From that point on radiologists were my friends.  That was an era before there was a lot of managed care penetration and I always rotated at public  hospitals and VA hospitals anyway.  You could always find a radiologist back in the dark confines of a reading room.  The interns and residents had certain staff members that were the go-to staff in terms of teaching and also amazing observations.  They always pointed out what we were missing.  They collected teaching files and teaching cases for us to learn from.  Reading rooms could be bizarre places in those days.  Very large films clamped on reading boxes.  In some cases entire rows of films - 10 to 12 wide, could be rotated on a belt device.  The radiologist would need to recall when they saw the film and press down on a foot pedal until the correct film popped up.  On many days row after row of films would need to be surveyed to find the one you wanted.  In the early days of spinal CT, many films had to be viewed on each patient.

I did not forget my positive experiences as a resident when I became an attending physician.  All the images I ordered on my patients had to be seen.  I would still go down and pull the films and where necessary review them with the radiologist.  Now I had neuroradiologists to work with and they were excellent.  The medium was changing.  Eventually all of the films went away and when I went down to radiology, the reading room was still there, but now it was a computer terminal with two monitors.  The images could be immediately manipulated to show the best view.  It was no longer necessary to pull the film off the cassette and illuminate it with a bright light.  I could always ask them questions, but as time went by they were under a greater time crunch.  Now all of the dictated reports were available on the phone system and you were encouraged to listen to all of the reports.  Asking to review a series of films without listening to that report was frowned upon.  At one point in time we were all members of the same clinic, but soon all of the radiologists were spun off into a different company.  They were the same people,  just no longer affiliated with our clinic.  By  that time managed care was trying to get everyone on a productivity scale and radiology seemed like an ideal speciality to crank up the productivity expectations.

In addition to the direct experience with radiologists, the author here also seems to not recognize the value of a human brain as a processor.  I teach neurobiology to students, residents, and physicians.  Part of the job of any lecturer is to help people stay awake.  Just before I delve into the frontal cortex and its connections to the ventral striatum, I put up a slide with a fact from one of my IEEE journals:

"Equivalent computing power (depends on the simulation) using today's hardware may require up to 1.5 gigawatts to power and that is equivalent to 0.1% of the US power grid or the output of a small nuclear power plant..."   IEEE Spectrum 2012

I ask the students to speculate on how the human brain has such a tremendous amount of processing power and how it is different from computers.  Even though the audience is generally tech savvy young physicians or students, I have never heard the correct answer.  One of the correct answers is the fact that the human brain is an unparalleled pattern matching device.  There are papers where it has been estimated we can each recognize about 80,000 unique patterns.  I start to go down the list and end with studies of radiologists, dermatologists and ophthalmologists demonstrating superior pattern matching and pattern completion skill.  But I also point out, it is why that you can't learn medicine from a textbook.  It is why you need clinical exposure before you can safely practice.   You need to acquire those skills.  To my knowledge, there have been no good papers written on available pattern matching in human diagnosticians compared with the cognitive tasks they face.  For example to be a good radiologist, how many unique patterns and variations do you need to be able to see - 10,000, 50,000?  The answer to that question is critical and yet we do not know the answer for radiology or any other medical specialty.  If the number if less than 80,000 (and we don't really know this confidence interval) - Watson may have the speed but not necessarily the accuracy.  Will Watson be analogous to the current ECG computer - a general normal/abnormal reading, a reading on measurable dimensions, and then not much on equivocal cases?  Only time will tell.

So I think this Health Care Blog post has the valuable lessons of most of their posts.  I don't know the author, but it is clear that he has not worked with radiologists as long as I have.  Not just the consultations backlit by reading boxes, but the telephone conversations about the best possible test to use to investigate the problem.  If he had worked with radiologists he would know that they have always been connected throughout the careers of most physicians.  The only obstacle to that connection has been corporate medicine.  The author's seemingly friendly advice is disingenuous.  If the business administrators who run health care really wanted radiologists connecting - they would get reasonable productivity compensation for that activity.  They would not need to connect and then run back to their terminal and read enough films to make up for the period of time they were in a conference or informally teaching residents from other specialities.  I think that the admonition to connect probably means to connect with the business administrators running the health plans.  Come back into the herd and let us tell you how many images to read, just like we tell other physicians how many patients they have to see.  Advising physicians on how to behave is also a well known strategy to manipulate them.

The real message is come back to the herd or be replaced, because there is nothing that would make an administrator's day more than replacing physicians with machines - especially physicians that they have no direct control over.

IBM knows that and I know that........

An equally important question is why Watson can't replace business administrators?  They seem to have the requisite lack of technical expertise and creativity.  They need a very basic level of pattern matching to do the job, certainly no training in it.  It would seem that a very basic program to optimize the working environment for physicians, health care workers and patients would be more ideal than dabbling in an area where real expertise and collegiality is required.  I can only conclude those concepts are alien to the ever expanding group of administrators whose reason for existence seems to be managing people - whether they need it or not.


George Dawson, MD, DFAPA


Supplementary:  Although I could not work it into the above post another insidious effect of corporations on medicine has been taking teaching out of the loop.  Radiology teaching files and teaching rounds were always a rich source of learning for students and residents.  It is a required skill on most board exams.  I recall approaching an administrator about preparing teaching slides for the residency in-training exam.  It is quite easy to copy de-identified images onto PowerPoint slides for review and these images routinely appear in all major medical journals.  I will never forget the response:

"Dr. Dawson - why would we want our images to appear on teaching slides?"

Just another sign of the apocalypse.










      

Physicians Replaced By Computers - Lessons From A Roomba

My Memorial Day project was purchasing a Roomba and getting it up and running.  I am a big believer that robots will make all of our lives easier at some point and decided now is the time to start walking the walk.  For those not familiar with the Roomba, it is designed as a robotic vacuum cleaner.  Once you have set up the rooms and programmed it, it is basically supposed to vacuum your floors automatically and then park itself in a docking station for charging.  The machine itself is about a 14 inch diameter disk that rises to a height of about 3 1/2 inches off the floor.  It is a light 8.4 pounds.  It is able to accommodate sharp angles with a secondary brush that spins on an arm that extends from under the main disk.  This combination of the main disk spinning and the extended spinning brush cleans the corners of a room.  I purchased the latest model, a Roomba 880 after consulting with friends and relatives who had earlier models.

One of the considerations in buying the Roomba was whether it would help turn my home into an even cleaner environment than it currently is.  That is a tough act.  One of my friends who is a physician gave his opinion that my home is "museum-like".  My office is probably the only problematic room with stacks of books and journals piled everywhere.  Disarray certainly but minimal dust.   My entire first level is hardwood flooring that is typically vacuumed with a built in system.  To its credit the Roomba contains all of the debris in the machine until it is emptied and all of the exhausted air is HEPA filtered to avoid exhausting any dust particles.  The main cleaning mechanism consists of two debris extractors that are rubberized bars that spin at a high rate of speed across the floor surface to capture dust, hair and larger particles.  But the most interesting aspect of the Roomba was going to be its observed behavior.  It has two modes when vacuuming.  It can start in a spot and spin increasing circles in an outward direction until gets to about a 3 foot diameter and then it spirals back in to the center spot.  In the more typical mode it heads to the room perimeter and then "automatically calculates the room size and cleaning time."  The most valuable tip in the manual was to take measures to restrict it to one room at a time and it comes with two Virtual Wall^®Lighthouse^TM devices that allow for easy demarcation of the work area.

The most fascinating aspect of getting started with the Roomba was going to be setting it up and watching how it went about the task of vacuuming.  I did some very minor room preparation, charged it up, and turned it loose.  As expected it headed straight for a wall and then attempted to establish the perimeter.  I remembered this as standard rodent behavior.  If you have ever confronted a mouse in an open area of your floor, their first move is to dash to the baseboards and run parallel to them to escape.  That strategy works well in the wild because the maneuver is associated with more cover and makes them less susceptible to predators.  It works much less well when confronted by a human who knows that it is their first move.  And yes, scientists have bred mice that do not exhibit this behavior.  My guess is that they would not fare well if they made it outside the lab.  The Roomba's behavior is less rigid than a typical mouse with some exceptions.  In the hour and 20 minutes it took to vacuum the adjacent kitchen and great room - it circled a kitchen island perfectly at least 10 times, but at the wall perimeters it was much less predictable.  At times the Roomba would peel off and take off across the room in a single pass or rarely return and continue along the original wall.  Sometimes it would head off the wall at a 45 degree angle and at other times 90 degrees.  There were never the usual adjacent passes that a human would make using a standard vacuum cleaner.

According to the literature,  the Roomba is supposed to "crisscross" the room in order to clean the floor.   I placed two small pieces of popcorn in the middle of a large section for flooring to use as markers of cleaning efficiency.  In the course of an hour, the Roomba passed these markers many times, sometimes very closely without vacuuming them up.  During that time it was very difficult to detect how much crisscrossing had occurred since mouse-like it spent the majority of the time in the periphery, bumping and spinning around walls and furniture.  It eventually did break free from the walls and set off on a 45 degree path picking up one popcorn fragment at about the one hour mark and the other at about one hour and ten minutes.  The old adage about pictures is true and I happened across this 30 minute time lapse photo of the Roomba working a room (with permission from the SIGNALTHEORIST web site).  It correlates well with my description of the actual paths.

As I surveyed the job afterwards, the floor was definitely clean and the warning light on the machine was saying that the dustbin was full.  When emptied, it contained an impressive amount of debris and dust relative to what seemed visible to the naked eye.  Another win for the robots?  Well, not really.  It is an interesting tool that I will continue to use and study, but in comparison with humans it is not efficient and at this point certainly not autonomous.  Despite all of the guidance in the manual the Roomba can still encounter unpredictable surfaces and get itself into trouble.  In my case it was the pedestal of a recliner.  The wood at the edge was about 3/4 of an inch high.  For some reason, the machine did not recognize it as an obstacle and continued to run up the base and get hung up.  A loud spoken error message would sound advising the human in the room to pick it up and start it in a new place.   The other concern is efficiency.  It spent far too much time in the perimeter and a low percentage of time covering the main floor areas.  That was tremendously inefficient.  It took at least 5 times longer to vacuum the main rooms than I would have if I was pushing a vacuum cleaner.  Even though it allows me to do other things, it says something about current state of available and affordable domestic robots.  They can't match the performance of humans on a fairly basic task.  This is an important concept.

Before any of the futurists out there jump on me for being a Luddite, let me disclose a few details.  I am a member of the IEEE and have been for the past 18 years.  I am currently a member of the IEEE Robotics and Automation Society.  I am not an electrical engineer and I have not designed or built any robots, but hope to start doing this when I retire from psychiatry.  I consider myself to be an expert in the human brain and the advantages it confers on humans over other animals and machines.  The Roomba is a basic case in point.  It cannot sense and adapt to novel conditions quickly enough to match a human doing the same task.  Even more striking is that although it is designed to vacuum homes and I have a fairly typical home with a better than average floor surface, it still encounters situations that exceed its response capacity.  In those situations it needs a human assist.  What is it about the human brain that leads to that kind of an advantage?  First and foremost, it is a rate of pattern matching and pattern completion capacity that allows us to recognize vacuuming problems, anticipate them and correct them by developing novel solutions even before the problem leads to a stop in action.  Some of this happens when a human goes around the room to set it up for the first time for the Roomba.  That human has made some assessment of the machines capacities and limitations and is problem solving for the machine before it is turned on.

Observing the limitation of the Roomba leads me to a point where I can address both the idea of computers replacing doctors and how that fits into the common anti-physician narrative in this country.  Is there a connection between the two?  My experience tells me that there is.  For nearly 30 years there has been a constant stream of antiphysician rhetoric.  The sources have been expected.  One of them is the key opinion leaders (KOLs) of the managed care industry.  I can recall reading one of the the first books written by one of them, a non-physician who was widely acclaimed as being an expert in managed care.  His early theory was that the high cost of health care was due to the decisions that physicians make.  But in the middle of the book he wrote what he thought of physician salaries and only grudgingly acknowledged that they should probably be paid a good wage due to their education.   I have posted here many times my experience at a managed care conference in the 1990s.  The speakers at that conference were very clear that the explicit agenda of their industry was to replace all of the specialists with primary care physicians.  The examples given were orthopedic surgeons and psychiatrists.  When a psychiatrist in the crowd pointed out the shortcomings of that philosophy - he was called a "whiner" by a Governor who was an anointed KOL in the industry.  Then the KOLs from the financial services industry started weighing in.  You could find glimpses of it while reading the investment literature.  People who were investors with no particular degree started saying that some day, physicians would get what they deserve - with the implication being that whatever that was - it was not good.  Any physician has experienced this prejudice.  The comments about how physicians are "expensive" as a rationalization for working them to death by not hiring any additional help.  Replacing physicians with computers seems like a logical extension of this rhetoric.  Googling this topic returns a number of provocative articles written from a point of view that is generally consistent with who the author is.

I know that some of those authors know the difference between a robot or a computer and a doctor, but it is also clear that some do not.  They certainly don't seem to understand that the real processing power of a human diagnostician's brain is in the area of pattern matching.  In order to duplicate that property with current technology, takes a massive computer and it is one of the reasons why my new $700 Roomba, although well designed - can easily be beaten by a human with a standard vacuum cleaner.  But the human advantage goes far beyond that.  Human diagnosticians do far more than match simple patterns.  They are able to complete fragments of patterns and anticipate what the whole pattern should be.  For example, is it likely that a depressed person is in this current state as the result of an inherited form of depression, their current state of detox from an opioid and/or benzodiazepine, current stressors or interpersonal conflicts, brain trauma, an undiagnosed medical condition, childhood adversity, psychological trauma as an adult, or defects in reasoning at either the emotional or cognitive levels.  Then there is the matter of acquiring all of the data to make the determinations.  Patterns upon patterns of data.  The Roomba-like approach would be to give the person a checklist of depressive symptoms and pretend that is all that needs to be known.  Checklists are already being administered by a computer and may be administered by robots someday.  

Yet it takes the pattern recognition, and several layers of it, as well as human experience dependent learning in order to make a real medical or psychiatric diagnosis.
        

George Dawson, MD, DFAPA






Supplementary 1:  The graphic at the top of this post is a photo that I shot of the inside of the box that my Roomba came in.

Supplementary 2:  I don't want to give the false impression that I do a lot of vacuuming.  My wife does practically all of it, but I am trying to do more especially if there is a high tech twist to it.  Some of the first robotics I hope to work on will be human controlled arms and hands designed to do yard work and move heavy objects around in the house.  I can't believe this is an area that has been ignored.