Fighting the imprecision in medicine using AI

In our current healthcare system, we mostly, we use an approach that’s called a one-size-fits-all approach, meaning if you are suffer from… I’m a dermatologist, so allow me to use some dermatology example. If you suffer from acne, ok pimples on your face. and you’re an adult typically the first line of drug will be this … the second line would be drug of this and the third line would be this… it’s all like standardize. It doesn’t matter, you know, how old you are male, female, you know, do you have a osteoporosis? or even do you have a diabetes or not, you know, the treatment are more or less the same.

That type of one side speak or approach has had served the medical community or the patient, you know, in the past few hundred years. but now we understand that this approach actually produced medical errors poor and inconsistent quality; waste and caused problems; also the ignorance of prevention. Because we were all focusing on treating patient but not preventing it. So what are… how bad is the imprecision in medicine? you know, diagnostic accuracy has been recently, well, only this year, diagnostic error has been cited as the number one patient safety problem. And when I reviewed the literature, you know, 20% of the diagnosis were incorrect and that’s a conservative estimate. Some people estimated 40% of the diagnosis as incorrect.

And so it’s a huge problem that we need to take home. Right now. Another example even there are so many examples about the imprecision in medicine. Although we, we talk about precision medicine like it the current practice but it is not. The current practice of medicine is still very imprecise, meaning we’re doing in precision medicine every day. Let me give you just one example to think about, you know, there are just so many examples but let me give you one example to think aboutㄡ Lab data range, you know so if we look at if we’re looking at liver function. Okay. One of the liver function is called AOT or AST, there are two.

So and the cutoff value for the AST is 40. So if you’re 39 you’re safe. Your liver is totally fine. If you’re 41, oh, so that’s in red, then you have to be careful because your liver may have a problem. And that cutoff of 40, there is no difference between 10 year-old girl anda 80 year old man. You know, they all use the same cutoff range of 40. Okay um And it’s not hard to imagine that the liver condition of a ten-year-old would be quite different than eighty year old man. Not to mention that the amount of alcohol and other… you know, drugs that that people are taking, will strongly influence the function of the liver.

But again, because of this one-size-fits-all approach, we use one simple range for almost all the 20,000 I mean 8 thousand that has that we can do most of them were just one normal range. and an order decision are made based on this one single range How precise is that? And if you look at allergy information that we did a study and we look at Energy Information, 40 percent of these information are are missing. Right? So we don’t even know. Whether the patient is allergic to something or not.

In family history it’s probably 90% incomplete, not to mention when we’re doing a diagnosis when we’re having a patient encounter the genomic the behavior in the environmental data, these are now available to doctors. So when you look at all this imprecision and missing data, no wonder the diagnostic accuracy is not very good, right? No wonder we got it wrong twenty percent or 40 percent of the time. So there’s a human diagnosis project that’s aim to actually improve the accuracy of diagnosis． and so when we’re looking at, you know, patient data usually this is a model that I use. We use five major type of data patient profile, diagnosis, problem list, procedures, medications, Lab examine.

So these are the five types of data that we use but then if you look at the big picture, these five type of data is only the phenotype. Okay, part of the phenotype. What we need more will be the genotype, the exposure type meaning it’s the environmental factors and also the behaviors and we need all these data from birth to now. So there is a temporal dimension involved. So we need… it’s not only that we need all these data, in order to determine the real health condition of a person we need all the data from the day that a person what’s wrong.

And with all this data we are able to do if we can if we know the diagnosis before it happens, we call it prediction. If we know that very early we call the early detection if we know the diagnosis, you know, after it happened, then we could provide suggestions or recommendations. And these are the clinical events that AI could actually help in terms of, you know, diagnosis treatment rehabilitation prognosis and management and you can see papers addressing all these different clinical events all the time. Supposedly AI are going to take care of us with all these clinical events with early detection or in our suggestion a recommendation or treatment to the day that we die.

So factors that affect house, it’s not just the phenotype, right? We also got the genotype that SNP, micro RNA Epi-genomics, pharmaco-genomics the phenotype, we talked about this. Exposotype including food, air quality, nor is lighting you know lighting condition could change a person’s psychology or the severity of a depression. Chemicals, temperature, humidity, radiation, magnetic field, even biometrics, even the air pressure, atmosphere pressure. Behavior like exercise, sleep quality, Alcohol, Tobacco, psychosocial pressure, work for drink, another activity including social life, these are all factors that could affect health. However, our doctors including myself, we’re only focusing on phenotype in most of the time. Actually, according to some study seventy-five percent of the time the phenotype data are weren’t even complete; they’re incomplete.

And we have to make a diagnosis out of incomplete partial phenotype data. So AI really has to kick in and save us from this problem. Some of the possible solutions, I mean there are so many solutions, but some of the possible solutions for AI to save the day will be background lifetime data surveillance because they’re just so, you know, there’s so much data to process. That’s humanly impossible. If I, as a patient, bring all my, you know, genomic phenotype and all these exposotype data to my doctor, actually, I got a patient who gave me a USB Drive; give me a thumb drive.

And tell me I would you look at that that’s my ECG for the past 365 days and see what’s wrong with me. You know, so how can a person a human physician actually deal with that kind of data. It’s impossible. And we always and we really really need smart summarization and visualization of the huge amount of data that the patient possess. And also just-in-time diagnosis and treatment advice like IBM for oncology is doing part of the treatment advice. We also need disease prediction and in early detection, that’s why that so we can do prevention or early interventions.

And also we also need adverse event prediction and detection, you know, so, we could prevent patient safety issues, medical errors or early interventions. So, it looks like AI for now, you know, we started with AI is better than human, and then, men lose jobs, which is bad idea . We don’t want that, right? And then people talking about, you know how about human plus computer is greater than human, right? And we all like it and we call we call it augmented intelligence right? Instead of artificial we call it mental intelligence and people like it.

But when I look at this formula, you know in the middle of the night, I realized that if it’s a mathematical formula, we kind of, you know, cancel out both human and then computer will rule the world. Of course this is a joke, but they also get you to think about, you know, things in perspective. Taipei Medical University, we’re doing something called AIMHI or artificial intelligence for medicine and health innovation. If you’re interested, you could go to this web address, AMIHI.tw. But I have to apologize first, because more some of the information are in Chinese and some are in English, so it’ll be a bilingual, kind of, a website.

So in conclusion, I hope that I convince you that imprecision in medicine is a serious problem. That’s why we, you know, always callyou know, people to do precision medicine because we’re just so imprecise right now. And you know big data is going to push AI to the next level. Everybody could agree on that and that’s why we’re seriously collecting data right now, for our patient, for health care.

And you know after thinking about it, you have to be very careful to say something it’s the only solution to something. But I have I would say AI is the only cost-effective solution to the very seriously one-size-fit-all problem. I couldn’t think of a better or a more cost-effective solution. Then I have to stop the problem because the reason why we have the problem to begin with the one size control problem fit our problem is because we are limited in time and resource in processing power to begin with.

So I guess if we if we do not employ something like AI to help us process all this information, we will never get there; we’ll never get out of the one-size-fits-all problem and we’ll never achieve the so called precision medicine vision that we talked so much about. Right? So, okay, my final sentence would be AI has to change the future of medicine or we may not have one and why is that? because we deserve it I hope so. That would conclude my talk. Thank you.