Here I will start to the second part of the tutorial. So I will focus on that. How to define the cohorts and how to characterize a cohorts and how OHDSI tools work. So I hope this class would be the most interesting class today. I need to try my best. So what kind of data evidence does OHDSI seek to generate from observational data? So we are trying to see evidence of three types. Clinical characterization, population-level estimation and patient-level prediction. So for clinical characterization, it’s kind of the assessment of the quality of the clinical pathways or and the natural history of the disease. So who are the patients? Who has diabetes? Among the patients with diabetes who takes metformin or sulfonylurea?
Who takes the…. how many clinicians follows the physical guidelines? Those are can be a quality improvement or the natural history of clinical characterizations. The second part. Okay. So the second part is a population level estimation. So it’s most more likely to about traditional retrospective research. So… such as safety surveillance or comparative effectiveness. So the example of the question could be does metformin cause hypoglycemia or does metformin cause hypoglycemia more than glyburide? Those kinds of things that we are… we define these kinds of evidence to population-level estimation And Patient-level prediction is actually the artificial intelligence theme. So you… I think most of you are interested in this part. So in this part there are personal medicine part or disease intercept.
So given everything you know about me and my medical history, if I started to take a glyburide, what is that chance I’m going to have hypoglycemia during the first days after starting the medications? Those kinds of the question you can find in the section of Patient-level prediction. So for those research, we need to define cohort first. Actually most of the medical research, we need to define the court first. So in OHDSI, we define cohort as a set of persons who satisfy one or more inclusion criteria for a duration of time. So there are several objectives based on this cohort definition. So, one person may belong to multiple cohorts.
And one person may belong to the same cohort at multiple different time periods. And one person may not belong to the same cohorts multiple times during the same period of time. Actually, it’s very confusing but you need to know that chords can be defined by the person and the time. So without that time, we didn’t define the court. It is always be with the time.
So I focus on each one by one evidences and show how we can do this. So for the natural history, if the question is that who are the patient having developed diabetes and among these patients who takes metformin? For asking this to this question, the target cohort should be the patient with diabetes and that outcome cohort should be patients start to use metformin. Does it make sense? Do you have any questions about this? So if you have… if you can define the target and outcome cohort, you can easily calculate how many patients start to use metformin among the patients with diabetes. And I will show how you can see the result of this question by using ATLAS.
For the quality improvement, if the question is the proportion of patients with diabetes experience, these are related complications. So for asking this, to this question, the target cohort should be patient with diabetes and the outcome cohort should be the patient developed complications. Does it make sense? Do you have any comments or questions about this? Okay. Again we figured… if you can define those two cohorts, it’s very easy to calculate the proportion of the how many people developing complications from the diabetes.
So let’s move to the second part. So population-level estimation. So first of all, if you want to know the metformin… whether metformin causes hyperuricemia the target cohort should be patient with diabetes. And the outcome cohort should be patient developing hypoglycemia. So if you can define or you can extract those two cohorts of it from your databases, you can easily find the evidence for this question. For comparative effectiveness, we need another cohort which the name is “comparator cohort.” So if you want to ask the… to the question does metformin cause hypoglycemia more than glyburide?
We need to set the target cohort as diabetic patients with using metformin and you need to set computer court as DM patients using glyburide and the outcome cohort should be the patients developing hypoglycemia. So you can… if you can compare the proportion of that hyperglycemia in the, between target and the comparatives, you can see which drug is better or worse. So it’s very easy
so again, more patient-level prediction or artificial intelligence, if you wanna see… if you’re on a predict the risk of diabetes or hyperglycemia in diabetes, the target cohort should be diabetic patients using glyburide and the outcome cohort should be the patients developing hyperglycemia. So based on this cohort system, you can easily build predictive model to predict hypoglycemia among the patients but using glyburide. So I will show how you can do this so easily. Again if you if your set target cohort as general population and if you set the outcome cohort as a patient of developing diabetes, then you can easily develop the predictor model for prediction of the diabetes among the general population
So I will try to show how the three the read(text) questions can be answered by ATLAS. So, first of all, I want to show how the estimate the proportion of a patient of a diabetes experienced diseases related to complications. And the second part, I want to show how you can compare the risk of the hyperglycemia between metformin and glyburide. And the third one, I want to show you how you can build the predictive model for predicting hypoglycemia among the patients using glyburide.