simtk.org

It was noted in the recent meeting that there are already computers in the clinic -- they are used for EMR primarily and for PubMed search secondarily.

Jacob noted in an email to me: "So there are models there already - what is the main obstacle for more models. And you mentioned EMR a point that is now of interest to me since NLP technology may be very useful to address these."

I'm not actually sure I know what Jacob means here but here are a few observations.

I suppose in the broadest sense the EMR is itself a model of the patient insofar as a database can be viewed as a declarative language model of something. Methods have been developed for "running" such declarative models to reach biological/pathological conclusions -- cf http://www.springerreference.com/docs/e ... 49798.html)

Of course, if you ask an MD in clinic about the presence of models or the use of models, implicit or explicit, he or she will be perplexed and will quickly change the subject.

Another related topic with respect to EMR is how difficult it has been to get adapted clinically and the unintended consequences of using EMR forms. Ironically and sadly, the medical record has become less reliable, albeit more available. It is not reliable since the doc simply checks off the tic boxes rapidly without thinking much or focusing on any individual item -- eg they will have check 2 reflexes and then check off the box for 'all reflexes normal'.

In reading EMR notes I often skip all of the tic boxes as unreliable and focus only on the freehand notes. In this respect NLP could be valuable in that, coupled with dictation (also currently unavailable except via human dictation services or via software in specialized areas -- notably radiology) it will improve validity.

This is an interesting topic indeed. Should one consider EMR as a digital model of a patient or simply as a digital dataset collected on that patient? If we rely on the latter description, I foresee a computational model more like parsing and analyzing that data to bring it in a more usable form. In a sense, it feels like a clinician who views EMR may not have access to a computational model (yet have access the raw data); the heuristic model in their mind (acquired through experience, scientific escapades, etc.) help them to cluster/classify the patient, and recognize patterns.

Cheers,

Ahmet

Hi Bill, Hi Ahmet,

Yes, this is a very interesting subject, and thanks for starting the thread.

First to Ahmet's point - EMR is data rather than a model. However, processing such data may be considered modeling. Modeling implies to me some sort of function or computational effort and although EMR represents knowledge it is not a processing/computational mechanism.

And now to the points Bill raised. Bill identified an important issue we must address in this committee - Technology Adoption. It is actually the most difficult barrier for the use of computational models.

There are two elements: 1) Novelty and 2) Human factors

1. Novelty - a new system changes things for the human and humans who are used to something and need to change their habits sometimes resist to this change. So new models will always have some adoption resistance.
2. Human factors - humans are not always good at repetitive jobs and systems that need complete knowledge should offer shortcuts for the human. This is at the level of the man machine interface. A good user interface may help the user, in this case a doctor, to quickly complete the task needed.

It seems that Bill is raising a User interface problem that can be fixed if the programmer who wrote the code would have to experience what the doctor has to go though - if only for a week. Data entry can be made easier by using defaults, similarities, and suggestive graphics.

I must admit, I had the same experience for about two months where I was looking at clinical trial database dictionaries. It took me about two months to extract only a subset of representative questions from the dictionaries of about a dozen trials. The work was demanding, repetitive, and generally not nice – although I learned a lot while doing it. The amount of information describing the patient today is staggering. Collecting all this data takes time and even going through it is becoming a non human task. The clinician is trained in collecting this information in his mind and analyzing it quickly. To let the computer assist in this analysis task we have to transfer the data to the machine. Letting the clinician do this task is perhaps a necessary evil – since the clinician knows the language the computer has to learn. Yet we should make this process as easy as possible to the human to avoid issues with data entry.

Bill suggested free text and hinted about speech recognition as an alternative data entry option to the machine. If this method is to be used, then NLP technology is essential to process the information. Never the less, there is already resistance to this technology - see my point about novelty.

I will be very happy to elaborate on resistance upon request – since this is an issue I pointed out in another forum, yet to keep this discussion relatively short I will do this only upon request. I do hope someone is curious enough to request further information about resistance to technology adoption.

Jacob

>programmer who wrote the code would have to experience what the >doctor has to go though - if only for a week. Data entry can be made >easier by using defaults, similarities, and suggestive graphics.

I have some 2nd hand acquaintance with the people who work at Epic, one of the major firms in the EMR field. They certainly place people in clinic settings as well as in the specific clinic where one of their products is to be placed. They also hire MDs to give talks to the programmers. Despite this, the problems persist.

I would note that most of us in the field have had some experience with getting comp sci or physics PhDs and getting them acclimated to the biosphere (ie modeling of biology) -- some get it in 6 mos, some in 1 year, some never -- i would say that none of them, with no prior experience, would get it in 2 weeks.

Anyway, I just have made this point to emphasize that this "cultural incommensurability" is not a trivial problem -- if anything the MD world is harder to penetrate than the bio-research world.

bill

Hi Bill,

Is the issue user satisfaction? Or is it something else?

Specifically to the problem you described in details, I can explain that the computer needs complete data in something that look like a table that should be filled completely. Even clicking through a big table takes time for the user and therefore may be annoying to the intelligent human forced to do repetitive work. This annoyance when accumulated eventually results in a dissatisfied user.

So there cannot be full user satisfaction here unless you can convey the information in other means that will fill the table quickly and completely. Unless the computer will have a model who can work with partial information.

If the issue is user satisfaction, then there are ways to handle it.

However, if there is something else then please do try to describe it. Perhaps by giving another example - if applicable.

You are touching at the heart of the issues this committee was built to address - this is very useful information.

Jacob

reposting? from aug 2

It is easy, all too easy, to fill the table quickly and completely -- although there is often not sufficient time to properly do all of the things that you are checking off on the forms (note that most MDs make an effort to at least do something for each system checked off so that they are not defrauding). A further downside of doing the detailed table is that one then loses the forest for the trees and is no longer pursuing a reasoned (and reasonable) diagnostic decision tree. This then could/should/and will be made into an engineering problem -- allowing the system (MSM or AI or Watson++) to formulate that decision tree and guide the clinician through it.

Right now there are at least 3 forces pushing the MD to perform more thorough medical exams: 1. EMR with its endless checkboxes; the others are stronger motivators: 2. you get paid more by checking off more checkboxes whether or not they have relevance; 3. you reduce your chance of being sued.

Getting all of this automated is the a mission of Eric Topol: {\em Creative Destruction of American Medicine} whom we're prob mostly familiar with? (if not I recommend checking out his posts at medscape.com since they are a shorter version of his book) -- he is an MD who is in favor of replacing MDs with machines ASAP. Of course his perspective generally raises the ire of other clinicians who look back fondly to the heyday of the traditional medical approach -- when reasoning rather than clerical competence was the key to diagnosis -- reasoning to figure out what needs to be done, while ignoring extraneous (checkbox) information.

Hi Bill,

You are pointing to a human dilemma. it seems that the clinician has many constraints imposed on them. And their opinions on what is the best path to take are split - you described a spectrum of voices.

In such environment human perception and human feelings play a role just as cold reasoning. This committee should be aware of these human factors, while making recommendations to help guide practice towards better models.

Specifically to EMR, NLP seems a promising analysis solution. Here are only some obvious aspects to support this claim:
1. More knowledge can be accumulated and analyzed than by a single doctor
2. A human is still making the decision yet has support that can help avoid pitfalls by providing additional analysis.
3. Some data entry tasks can be automated - and perhaps even given to the patient or a device.

This committee should help with adopting such new technology by pointing out what is credible practice with this technology. There will always be a spectrum of voices and the committee should take this into account.

I hope we see this the same way.