A precision trial design project we did, where we were trying to emulate the trial and demonstrate how it could have been optimized. Identifying the patients to be in the trial. If we used the trial criteria that the client had, it required an understanding of and reading of the ECG. So first of all, if you only had structured data sets available, you wouldn't have been able to replicate and then optimize the trial because you just have no idea what their ECG would have said, because you don't have ECGs. You just have structured tables. So we were able to pull in the ECG and the cardiologist's interpretation of the ECG, which showed you what it said, and use that to define the patient cohort and define all of the other criteria on top of that so that you could actually more accurately replicate the trial cohort, the patients who were going to be in the trial. 

Then, when you're trying to optimize for your trial on top of this, you know that you're doing so from a place of accuracy. Like the types of patients you would have gotten if you didn't have access to something like the ECG, the physician's interpretation of the ECG, and the abstraction using a large language model on top of the physician's read of the ECG to pull out all of these interesting diagnoses and labels based on their interpretation of it. If you hadn't had access to that, you would have ended up with a patient population that looked far healthier than the population actually is in real life. And so you would have designed a trial based on that set of criteria, and that may have missed the mark in terms of whether or not you actually end up seeing the types of events that you're looking for to demonstrate that your intervention is successful in a trial. 

Because if you're working, if you're drawing from a pool of healthier people, then they're not going to have as many outcomes, their outcome rate for negative outcomes that you can impact is not going to be as high, and you may not be able to show that your intervention, your treatment, your therapeutic is actually as effective. So, phenotyping your patient population correctly really matters from that perspective.