Jan 26 2010

Digital Doctors

This concept may bring new meaning to the phrase “Doc in a box” (used to refer to small walk-in clinics). Increasingly computers are infiltrating the practice of medicine – but to what extent have or will computers replace the cognitive work of trained health professionals? This is a concept I have been following with interest, but at the moment there is probably much less in the way of computer-assisted medicine than the the public imagines.

A news story today reminded me of the baby-steps that are being taken in the direction of AI (artificial intelligence) medicine. A new study shows that a computer program is as good or better at making a specific measurement (wear of the meniscus) on MRIs of the knee – but only when they are mild to moderately damaged, severe damage still requires a human eye.

This kind of thing is definitely the low-hanging-fruit for medical AI systems – interpreting digital images. It does not require making a diagnosis, weighing choices, or interpreting human input. It is simply using pattern recognition to measure one feature of an image.

But even with this application, computers have a hard time competing with humans. People are very good at pattern recognition – something we still do far better than the most advanced computer. My own experience with this in medicine is in doing nerve conduction studies which measures the electrical response of an impulse sent through nerves to measure their function. The output is a simple waveform, and the computer program is tasked with suggesting where to mark the beginning, peak, and end of the curve. Sounds simple enough, an yet I almost always have to tweak the computer’s markings, and if there is any noise in the waveform the computer can be way off, sometimes even marking a bit of noise that is not even the nerve response.

What this program does well is making all the calculations – and for some types of studies even very complex calculations. But of course, that is a computer’s sweet spot. Pattern recognition – not so much.

What I find interesting is thinking about what computers do well vs what people do well, and therefore where and how best to integrate computer systems into the practice of medicine.

Simply handling large volumes of information (so-called medical informatics) is primarily the current role of computers in medicine. The best systems do this well, but in actuality they are only as good as the implementation – meaning the way in which a hospital or practice customizes, uses, maintains, and supports the application. For accessing and documenting medical records, lab results, viewing digital images, etc. computers are awesome and I would never go back.

But even here, the state-of-the-art of electronic medical record systems and their implementation is about a decade behind where it potentially could be. The reason for this, I think, is that implementation is very difficult, and requires high-level communication between medical experts and computer experts. Computer experts do not know what physicians and nurses really need, in terms of information and workflow. And medical experts do not necessarily know what computer systems can do for them, or how to communicate their needs in ways that translate to the software developers.

The process is slowly inching forward, mainly due to the efforts of computer savvy medical experts who can bridge the gap. But it is frustratingly slow.

What about reading digital images, like in the current news story? That is probably one of the next areas ripe for development. As I said, I don’t think computers will replace humans anytime soon for such readings, but they can provide an excellent assist. What computers do well (and humans do not do well) is provide consistent thorough attention. Every radiologists nightmare is having their attention waver for just a moment, and missing that small detail that has significant implications for the care of a patient. Perfect vigilance is hard. In fact all of medicine is plagued by the occasional lapse of vigilance and attention to detail, leading to medical mistakes. We have increasingly put systems into place to minimize such mistakes – but all we can do is minimize them, not eliminate them.

However, computer assistance can provide the vigilance and systematic attention to detail that humans find so challenging. I can envision the development of sophisticated algorithms, for example, that can systematically read X-rays and make precise measurements. The results can then be presented to a human radiologist, who can filter out the noise and false positives, and provide the context and pattern recognition that humans are good at. This way we get the best of both worlds – human and machine.

This type of approach is not new – such systems are called expert systems, as they provide an assist to experts but do not try to replace them. There are many areas of medicine where this can be helpful. For example: flagging abnormal lab results and bringing them to the attention of the ordering physician; checking for drug-drug interactions or allergies whenever a new drug is prescribed (assuming that the patient’s medication list is up to date); suggesting possible diagnoses from a list of signs and symptoms (maybe suggesting something the physician did not think of); or reminding physicians of standards of care (imagine this in a Hal-like computer voice – “your patient is over 55 and has hypertension and diabetes; would you like to prescribe an anti-platelet agent for vascular prophylaxis?”)

In short, there is tremendous potential for computers to function as assistants in the practice of medicine – reducing error, improving adherence to standards of care, and giving physicians access to the information they need when they need it. I do not think we are fully exploiting this potential, however, and while we are considering ways to improve the practice of medicine to reduce errors and improve cost effectiveness, this is an area that seems worthy of investment.

But we are nowhere near computers actually practicing medicine. For the foreseeable future the practice of medicine will be a partnership between person and machine – each doing what they do best.

15 responses so far

15 Responses to “Digital Doctors”

  1. EvanHarperon 26 Jan 2010 at 11:06 am

    Doc in a Box? Wasn’t that a skit on Saturday Night Live?

  2. Jerryon 26 Jan 2010 at 12:11 pm

    While people are good at pattern recognition, are we not sometimes too good at it? I can see a potential computer assisted xray technician seeing patterns that aren’t really there because the computer said there was something. ref. pareidolia.
    Might lead to many more fals positives?

    Possibly circumvented by first looking for things without assistence, then showing wha the computer thought was something interesting, and comparing.

    I don’t know why I still get amazed at how little use is made of technology that’s been around for years.

  3. calinthaluson 26 Jan 2010 at 12:41 pm

    In AI terms, IIRC, an expert system is actually defined as a non-, or minimally learning system. It is put in place by getting data from experts (doctors in this case) and being as good as it will get at the point of installation. This is opposed to a growing, learning system where the AI starts as a child and has to be fed data slowly to develop connections much as a human child. This bottom up approach is a neural network or sub-symbolic approach.

    It is not defined as expert because it aids experts…but because it is derived from data from experts. This is limited because many domain experts cannot codify their expertise in a practical way, and it has no gift for intuition or creativity. Also any additional knowledge requires an update to the base knowledge center…there’s no good way for it to advance or adapt to changes without significant alteration by a human.

    I’ll take pedantism for $500 Alex.

  4. superdaveon 26 Jan 2010 at 1:55 pm

    Finally a post on something in which I have some expertise.
    To really grasp this field one has to ask, what do computers do well?

    one, they have more higher and accurate memory capacity than humans

    two, they can do calculations faster.

    You add this together, and I think what this spells is a computer system in which the computer is used to analyzed a large database of patient data to do intial sorting of trivial cases. However, I don’t have enough medical knowledge to know if such a database of undiagnosed patients exists or would be useful.

    In order for this to really take off, the computer has to detect things that could not otherwise be seen by a doctor. How do you test this though? If the machine says your prostate has cancer and the doctor disagrees, the only way to know who is right is to wait or get a biopsy and both those options are not ideal for the patient.

  5. Roger Bigodon 26 Jan 2010 at 2:35 pm

    The idea of a human-computer “partnership” has been fruitfully applied to chess. Kasparov discusses it in an excellent piece in NYRB.


    It’s an interesting testing ground, since the scoring system has been subjected to a lot of statistical analysis and there’s an objective approximation of “ability”.

  6. sonicon 26 Jan 2010 at 3:23 pm

    I would think help reading x-rays could be a good use for computers-
    I know the doctors in the emergency room where I worked would have loved it if they could have a machine that asked-
    “Is this a problem?” (indicating some area of the x-ray) ((Especially at the end of a long shift…))

  7. BubbaRichon 26 Jan 2010 at 3:26 pm

    I went to a doc-in-the-box over Xmas holidays, when my doc was gone, with a fairly severe sinus/throat ailment. She seemed to type my symptoms and other answers to her questions into a diagnostic program, which suggested other questions. She prescribed a fairly strong coupla medicines. They didn’t fix the problem, though my doctor didn’t criticize her prescription at all, I just had a different bug or a resistant strain. I don’t feel that comfortable with a doc relying on that program for common ailments, although I’m happier if it also tries to exclude more serious possibilities.

  8. Draalon 26 Jan 2010 at 7:08 pm

    I wonder if the more integrated computers are in medicine, the poorer the quality of physician. Sort of like Wilde’s theory of Risk Homeostasis. If a doctor relies on computer assistance too much, than possibly the less effective the doctor becomes, especially if the computers are suddenly taken away. Take calculators for example: I first learned basic algebra by hand, became fairly adept at it, but when I was allowed to used a calculator, my ability to quickly do pen and paper calculations decreased dramatically. Another example I encountered a lot was when I was a Teaching Assistant of undergraduates who used computer programs to solve their math problems. As long as the computer spit out an answer, they assumed the correct answer was reached. Repeatedly, I even graded homework that was 20 pages of computer error messages.

  9. cobaltnineon 26 Jan 2010 at 9:02 pm

    Right now, I’d guess that some of the pharmacological uses of computers in medicine are among the lower-hanging fruit, even before image analysis. Perhaps due to the more limited focus that the pharmacy systems have, they are working more clearly – instead of trying to do everything at once. They also have the advantage – most patients only go to one pharmacy, but many will go to several specialists, and communication between them can be tricky.

    Given the updated med list (and I see plenty that aren’t), and things such as height, weight, and, say, recent blood-sugars, the pharmacy portion of medicine could definitely take more advantage of the power of computers to synthesize data according to predefined rules and present suggestions for providers to alter medication as needed. Even now, although very irregularly, sometimes in outpatient I’ll see a faxed notice, usually from the insurance-preferred mail-order pharmacies, suggesting a particular class of drug as an alternative, due to other medications which the patient is on. I’m not entirely sure about their motives, but if that systematic analysis could be harnessed in a neutral/medical setting, that would be nice.

    I work in an outpatient setting and cringe whenever I hear that the pharmacy caught a medication allergy/cross-reaction. At least someone caught it. We should have.

  10. zoe237on 26 Jan 2010 at 9:49 pm

    I was thinking of the chess match too- what’s the Harry Potter quote? Never trust anything that can think for itself if you can’t see it’s brain? Okay, so that was Voldemort’s diary, but still.

    Seriously, I was disappointed (in retrospective) for the $2000 MRI for my own meniscus/ACL tear because I’ve read since that good orthopedists are just as good as the MRIs at diagnosis. My orthopedic doctor called it right himself before any MRI. And he’s cheaper too.

    I’ve also read that computers can make drug/ diagnostic mistakes as well, and that this is a big source of medical errors.

  11. SimonWon 27 Jan 2010 at 2:09 pm

    Human’s have too much conceit, and are too error prone, to be trusted with the interpretation of challenging medical data.

    Give me an objectively trained statistical tested computer system every time.

    Almost all areas of skill I’ve looked at, where a computer can make a useful contribution it usually rapidly exceeds practitioners in the field. Chess is a good example. Whilst Kasparov might have been winning 32-0 in 1985 against computers, even then >90% of humans who play chess would probably lose to a computer. These days all but a handful of humans would be crushed.

    Quite often these tasks are done by humans long after computers can do them much more effectively, because humans are good at adapting the world to keep them gainfully employed, even when this isn’t in the interests of their clients/customers/employers/patients.

  12. Matlatzincaon 28 Jan 2010 at 7:07 pm

    Interesting discussion Steve, but I think you miss the mark on why EMRs are hard to implement. There are actually good instances of EMRs that take into account workflow requirements and clinical applications (both patient-care and research). However most organizations have legacy systems that can’t be ditched en-masse for new (often very expensive) cross-compatible systems, and you can forget about communication between institutions, even if they use the same EMR vendor.

    One of the only solutions that I can see moving forward for EMRs is to have a standard set of requirements such that all of the individual software programs communicate with each other. I very much doubt that the insurance companies will be able to get their act together enough to create such a standard, which means that it may have to come from the US Government.

  13. Elwoodon 28 Jan 2010 at 8:46 pm

    One of the only solutions that I can see moving forward for EMRs is to have a standard set of requirements such that all of the individual software programs communicate with each other

    This is still an issue in the area of digital medical images.

    The DICOM standard of transmitting and archiving medical images is something that all vendors use – but the DICOM standard does not include all possible pieces of DICOM information. So 2 vendors required to connect their products (e.g. a ultrasound machine from one vendor to an archive server of another) can both be ‘DICOM compliant’ but still have errors and failures in data transfer if one vendor is using different DICOM information to (for example) sort and filter images.

  14. Greg Goldmakheron 29 Jan 2010 at 9:52 am

    As a radiologist, I have worked with automated detection/recognition systems, and at least right now they are dismal. The area that they’ve had the most application so far is mammography, where there are several systems which do a “second look”, trying to detect signs of cancer that you might have missed. The problem is that there are many visual features that malignancies and “fake-outs” have in common, and distinguishing the real from the artifactual depends on factors like comparing multiple views, current studies with old studies, and so on.

    In general, visual recognition tasks are hard for automated systems, though they are getting better. We can make a computer that will beat any human being in chess, but a human is still much better at walking into a room and finding the chessboard.

    Doing simple measurements is fine for automated systems, and has its place. Making actual diagnoses, though? Humans may be error-prone and conceited, but if my mother, sister, or wife gets her mammogram, I want it read by a radiologist rather than any piece of software.

  15. […] I recently discussed elsewhere that there will likely be an increasing role of expert systems in the practice of medicine. This  include things like systems for analyzing radiographic studies and highlighting potential pathology, checking for drug-drug interactions when new prescriptions are written, suggesting possible diagnoses to be considered, and, yes, running through checklists or algorithms of proper evidence-based management. This may be as simple as reminding a physician to consider prescribing cardiovascular prophylaxis to their 60 year old patient with hypertension (something which does not happen as often as it should). […]

Trackback URI | Comments RSS

Leave a Reply

You must be logged in to post a comment.