8
$\begingroup$

This is in followup to a previous question. here :

Neural network model to predict treatment outcome

and might be considered to refer to a different aspect of this question:

Application of machine learning techniques in small sample clinical studies

Thanks to Zach who suggested reposting.

I've put some fairly serious reading in now on CART, randomForest, Neural Networks and machine learning in general, learned about WEKA and the R packages, seen and followed the Stanford engineering lectures http://www.ml-class.org/course/class/index , I'm 3 chapters into Hastie. Given the kind of data we see regularly in clinically oriented research - loads of clinical parameters + loads of biochemical parameters + pen and paper test data +/- neuroimaging data with smallish numbers, I get the feeling that I'm missing something. I'm not regularly reading about ML techniques being applied in the research literature. My question is: have I just latched on to something which is dubious and therefore regarded with justified suspicion by researching clinicians and biostatisticians who are well aware of it, or are these techniques genuinely overlooked or feared outside of "business analytics"? What keeps it "niche"?

$\endgroup$
1
  • $\begingroup$ I think the problem here is more related to what journals you read than anything else. Machine learning is quite heavily applied in modern translational medicine, provided that black-box models are acceptable for the task. $\endgroup$ Mar 4, 2015 at 15:38

2 Answers 2

7
$\begingroup$

Machine learning techniques often lack interpretability. Also, they tend to be rather crude from a statistical point of view --- e.g. neural networks make no assumptions about the input data. I have a feeling that lots of people (especially if they have a strong statistical background) look down on them.

$\endgroup$
4
  • $\begingroup$ Yeah, I think I got that. However, from my point of view statistical methods are neither pure nor dirty, just the application of logic to data. If you want a pill to cure something, then you need to understand the interrelationships and take that to the molecular biology lab. However, If you just want to make a prediction using black box (NN/RF) or decision (CART) methods, what's the problem? You might even gain insight. Is it any deeper than snobbery? $\endgroup$
    – user6666
    Dec 4, 2011 at 19:26
  • $\begingroup$ While interpretability is certainly nice, I'm not sure whether I would consult a doctor who knows what he's doing and has a success rate of 60% versus a doctor who has no clue but has success rate of 100% ;) $\endgroup$
    – blubb
    Dec 5, 2011 at 13:06
  • 1
    $\begingroup$ You might be interested in Leo Breiman's 'Statistical Modeling - The two cultures', where this is covered in depth(recognition.su/wiki/images/8/85/Breiman01stat-ml.pdf) Furthermore, there are reasons for this approach -- if you want humans to interprete things, for example. $\endgroup$
    – bayerj
    Dec 5, 2011 at 21:28
  • $\begingroup$ @blubb I would consult a doctor that has no clue but 100% success rate with a confidence interval [98,100] :) $\endgroup$
    – Simone
    Dec 11, 2012 at 10:39
4
$\begingroup$

The track record of machine learning in biomedicine has not been very good. Early successes in machine learning came in high signal:noise ratio pattern recognition areas such as visual pattern recognition. The S:N ratio is much lower in biology and the social sciences. Machine learning effectively fits a lot of interactions between predictors, and to do that you must either have a huge sample size or a very high S:N ratio. See Is Medicine Mesmerized by Machine Learning?. In addition, many practitioners of machine learning has misunderstood prediction tasks as classification tasks. See here for more.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.