Machine Learning in Medicine

Friday, April 24, 2009

Now, this blog is full of technology-related stuff. I’m also interested in medicine, so today I’m going to share some of what I’ve been thinking about.

People are complex creatures, and so medicine is a complex subject encompassing multiple fields of discipline. Many technologies have been developed to advance medicine to its current form, but one particular aspect of medical technology remains elusive: how to supplement physician’s medical knowledge using computers. Now, I know what you’re very likely thinking: when I catch a cold I want to see a doctor, not a machine! I too would much prefer to see a doctor. However, it must be noted that machines are already playing a significant role in medicine and they could propel medicine so much farther if we just knew how to use them properly.

Medical intelligence is a branch of artificial intelligence dedicated to using machines to help diagnose and treat diseases. Many doctors see dozens of patients every day, and must have amazing physical and mental stamina to serve each patient reliably. Wouldn’t it be nice for a doctor to have a PDA capable of suggesting courses of action, allowing her to focus on the more complex aspects of treatment without tiring as easily? Such a system would not only streamline the process of seeing patients, but also improve speed and accuracy. The question, then, is this: how do we build a computer system with intelligence comparable to a doctor’s?

The answer is simple: send it to school! This is how we learn as humans, so this is how our mechanical doctors must learn. But we don’t send it to just any school; we let it apprentice with real doctors all over the world, so that it may have the knowledge to treat the whole world of disease. As soon as it learns something new, the knowledge becomes instantly available to all doctors using it. Medical knowledge has never spread so quickly.

Current medical intelligence systems are inefficient for a few reasons:

  • given the vastness and complexity of medical knowledge, it is difficult to construct a single program to deal with all of it
  • because medical intelligence is fragmented, it could only be used in isolated situations where contact with other disciplines is limited
  • since the fragments are isolated, it is difficult to network them into a single entity
  • without a network, the fragments cannot share knowledge and improve themselves effectively

Therefore I suggest the following:

  • create a universal medical intelligence system using the vast distributed computing resources available today
  • allow anyone with medical knowledge to use the system
  • as the system is used, it improves itself
  • every user has access instantly to the latest knowledge that is incorporated

One of my current projects has been to set up a freely available system for building classifiers. I have limited computational resources, but if this takes off it may form the foundation for something greater…


  1. Boris says:

    I think setting such a thing up would be extremely difficult. I don’t think most doctors would like the idea of a machine making decisions for them. Machines can only know as much as is put in, and sometimes years of experience can be more useful…especially if there might also be ethical concerns at stake. Sometimes computers can also be wrong because they will never come up with new ideas.

    That being said, there are already some such machines used in practice though I don’t remember the details and I think their use is pretty limited for the reasons stated above. I do think machines can play an enormous role in medicine, but this is a long term goal and requires a lot of discussion. There need to be a standard code for how such machines will be put into practice efficiently, how they will be used, and how their output should be handled by healthcare professionals. It’s far easier to utilize machine learning algorithms in scientific studies with applications to medicine than in medicine itself.

  2. Jiang Yio says:

    No, a doctor would not like the idea of having machines making decisions for them, especially having spent so much effort to become doctors themselves. But — this is more about society not being ready for changes that may be beneficial to it. After all, nobody knew what to do with penicillin when it was discovered.

    Incidentally, I’m about to step out to my Premed Advisory Committee interview, so thanks for getting me to think about these things :)

    sometimes years of experience can be more useful

    Naturally, and human experience is limited to a lifetime whereas machine experience can arguably span multiple lifetimes (simultaneously, even).

    computers can also be wrong because they will never come up with new ideas

    The same could be said for mistakes due to new ideas… but at least the software inherits our new ideas as we use them, right?

  3. Shuhao says:


    We now have Watson.