back to article Not so fast AI Doctor, the FDA would like to check how good you really are at healthcare

The US Food and Drug Administration has proposed a framework on how it might regulate medical devices that rely on AI and machine learning algorithms. The report published this week outlines two types of algorithms for the purposes of regulation: “locked algorithms” and “adaptive algorithms.” Locked algorithms provide the …

  1. BebopWeBop Silver badge

    Fundamental disconnect

    On the face of things, the FDA recommendations make a great deal of sense. There is a disconnect however between many of the people developing these technologies who frequently rely on the mark 0 eyeball - does it appear to provide 'better' performance than what came before and those in the hard real-time development community who demand a little more than that - just because some addition to (say Boeings flight management software) looks superficially better and possibly for trivial reasons, does not remove the responsibility of proper analysis and testing. Difficult to achieve with many 'learning' techniques.

    I would hope the latter prevail.

    1. jmch Silver badge
      Thumb Up

      Re: Fundamental disconnect

      What certainly has to be enforced is the recertification after every change.

      1. BebopWeBop Silver badge

        Re: Fundamental disconnect

        Just like Boeing did (oh whoops)

        1. Trollslayer Silver badge
          Mushroom

          Re: Fundamental disconnect

          US aircraft manufacturers can do a lot of self certification.

          Oops.

    2. Korev Silver badge

      Re: Fundamental disconnect

      I'm involved on the periphery of clinical trials. We tend to distrust pretty much any readout that relies on human judgement; however often there is no alternative. For example if you're looking at a MS patient's movements and video someone touching their nose (one of the tests in clinical use) it's hard to get two humans to agree on how well a patient does it.

      1. EU time zones

        Re: Fundamental disconnect

        It's all bound up in the gigantic assumption that objective data is good and subjective data is bad. Which is the root of a lot of the high-tech hangup thinking. I was in the lab end of Clinical Trials, and we could honestly be sure about our objective, lab-data only because we ran lots of QC, we controlled our environments etc. All very CLIA (USA) and Clinical Pathology Accreditation Ltd (UK, RCPath). And then the Home Office, to save a few quid, outsourced all of the forensic labs work to some liars and incompetents who couldn't even prevent sample switching, let alone assure good quality. And we haven't begun to near the end of that clusterf**k.

        You *can* cope with human subjective interpretation, it's all well established science, BUT IT'S REALLY EXPENSIVE. At *least* two different observers have to asses the data themselves, provide their answer and then a third party has to compare the results. We used to do this for University Exam marking - any papers more than 1 grade apart from the two markings had to be re-marked.

        Would I have my diagnosis performed by an AI-machine. Not if I can help it, and will reach for the family lawyers if I find that it has been.

  2. ProfessorBlockchain

    Adaptive learning

    If an adaptive algorithm is one that adapts its training in real-time (versus a 'locked' algorithm that is trained using a fixed set of data prior to deployment) then I'm not sure it's correct to say deep learning algorithms are adaptive. Most deep learning networks are trained once and fixed for deployment. Very few if any deep learning algorithms can learn on the fly in this way because of the computational overhead (GPU time) needed to train or update the neural network.

    1. Medical Cynic

      Re: Adaptive learning

      The human version learns as it goes along, by its mistakes. Hence M&M [mortality and morbidity] meetings, MDTs [multi-disciplinary teams] and SUI [serious untoward incident] analysis.

      Will adaptive AI be able to do the same - meetings not required, and all the info is likely to be available in the patient record?

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2020