ARTIFICIAL INTELLIGENCE
( continued from page 15 ) as well that it took me less than a minute to perform this “ analysis .” Part of the danger with the familiar is that we are often lulled into forgetting the complicated underpinnings . The reality is that there is a long list of assumptions required to run regressions . An excerpt from “ The Statistical Minute ” section of Anesthesia and Analgesia reads , “ The residual are the differences between the observed values predicted by the regression model , and the residuals must be approximately normally distributed and have the same variance over the range of predicted values ; the residuals are also assumed to be uncorrelated . In simple language , the observations must be independent of each other ; for example , there must not be repeated measures within the same subjects ” and the list continues . 1 So a bit of machine learning has been with us for some time now , the question is whether this is more of the online chess game variety or the computer flying an airplane sort .
Figure 4
Let ’ s swing the pendulum back to the other side and examine some deep learning literature . In a recent manuscript in Gastroenterology , the authors reference publications that utilized deep learning in capsule endoscopy ( Figure 4 ). 2 The surprising thing to me wasn ’ t the number of articles they cite , but rather the technology each used . Except for the SSAE ( what this represents is outside the scope of this discussion ), each one used is a convolutional neural network . Juxtapose this information with the regression anecdote above , and consider : when was the last time you heard of a clinical researcher piddling around with a convolutional neural network ? Not only is running these analyses not nearly as straightforward , these complex neural networks have been meticulously created by top computer scientist teams . This is to say that it ’ s not the neural network that is generated anew each time a new project is carried out ( usually ), instead researchers simply change the data they input into the system . I cannot give you an example of a neural network output , because I lack the skills to even generate an output .
The point of the last two paragraphs is to say that artificial intelligence in its primitive form has been in use for some time already , and not only that , but it is being reported routinely in the literature . It is being used already to make clinical decisions regarding patient management . Alternatively , the use of some forms of machine learning , which has caused so much angst , still requires a fair amount of sophistication to carry out . Further , it relies on pre-existing models created by highly trained computer scientists . Considering that few machine learning algorithms are in broad use clinically to date , we must accept the fact that it is more likely an errant chi-squared , ANOVA or regression has negatively affected patient care . It is more likely these basic analyses are more likely to
have negatively affected a patient so far you might say . The future is another issue .
I had previously assumed that the black box of computer programming was larger than it is . The term for when a computer really thinks for itself is termed unsupervised learning . The alternative , supervised machine learning , requires humans to label data ( Figure 5 ). That is to say that a highly trained computer person has created a very specialized piece of code , and typically another very highly trained individual labels lots and lots of data so that the machine can learn what it ’ s looking at before being asked to make a judgement on a separate entity it ’ s never seen . My favorite example of unlabeled data being fed into an unsupervised learning algorithm is the story of when Google showed 10 million video thumbnails to an untrained computer and it began searching for videos of cats .
If there is one statement to stress it is the one above . When browsing manuscripts , the first place I usually start is to ensure that the authors are describing a supervised learning model . Except in scenarios where little is known and we are only trying to search for patterns so that we can be pointed in a general direction for future investigations , unsupervised learning has no role in medicine in my opinion . Now it ’ s time for a disclaimer : I do not have a degree
16 LOUISVILLE MEDICINE