So, of course, the topic of today “Artificial Intelligence in Quality and Patient Safety” is an important one and an exciting one especially because Artificial Intelligence is really moving forward at a very rapid pace nowadays. And things as mundane as helping you pick your next Netflix movie,
filter your inbox for spam, all of those impressive voices systems that we now have today, let’s not forget self-driving cars. All of this is very impressive, and making us think what, for me, was maybe the most impressive thing this year on AI was this presentation by the the CEO of Nvidia at the Consumer Electronics Show where from his pocket, he pulled out this chip that has a supercomputer on it, and so this this chip is actually specifically designed to run deep learning neural networks. And so we’ve come a long way from not just developing this software, we’re now actually getting the hardware that it is dedicated to running that software.
And so I thought, well, let’s think about how much AI is influencing and happening in my field of expertise which is learning about the safety of drugs. So how does that work? So I picked a paper more or less at random, and so this paper studies the effect of a drug isotretinoin on a specific outcome, inflammatory bowel disease. And it uses a health insurance database in many ways, similar to the one that you have here in Taiwan, but this one is in the US. And so from that large databases selected 8,000 cases of inflammatory bowel disease, 21,000 controls. They selected, extracted seven variables for those people, and they ran a logistic regression model.
Just idea over the magnitude of this computation, they have a matrix of 30,000 rows and seven columns and they fitted a logistic regression. I could have run that on this computer from the 80s. I don’t need the fancy supercomputer on a chip for that. So I’m a bit frustrated by the state that we’re in when it comes to studying the safety of drugs and of treatments in general. And so what I would like to go into today with you is, well, how do we get this field into the 21st century. And so I’ve described three easy steps that I would like to take you through. Step one is one that I think Dr.
Westbrook would agree with me is we need to measure performance. We need to know how well this system works. So when I started in this field, that was actually a little bit, people thought I was a little bit odd like, “what do you mean?” “It’s a study, it just produces an answer.” That’s the answer. But I would say, no, this is, this study is basically a measuring device. We have a design that we run on this study design, that we run on this database. It produces an answer. And that answer that estimate of the relative risk or the odds ratio in this case.
It’s just like when we use a scale to measure weights, we can measure the accuracy of that instrument. So if we want to measure the accuracy of a scale, we put weights on it where we know what the true weight is a gold standard. That’s exactly where the word “gold standard” comes from. A weight where we know exactly how much, how heavy it is, we put it on a scale, and we see whether the weight, the scale actually points to the right weight. We have a similar thing that we can do with these observational studies. We can use control questions, so these are simply exposure outcome pairs where we know what the answer is, what the odds ratio is.
In general, these are actually a little bit more tricky than I’m making them sound, so my favorite ones are called “negative controls,” where the exposure outcome pairs are simply believed do not have a closed relationship, where we do not believe that the exposure causes the outcome. And we’re, therefore, the odds ratio should be one.