
Patrick Leung
Hong Kong filmmaker
3
APPEARANCES
3
PODCASTS
012
DEC 30
JAN 6
JAN 13
JAN 20
JAN 27
FEB 3
FEB 10
FEB 17
FEB 24
MAR 3
MAR 10
MAR 17
MAR 24
MAR 31
APR 7
APR 14
APR 21
APR 28
MAY 5
MAY 12
MAY 19
MAY 26
JUN 2
JUN 9
JUN 16
JUN 23
JUN 30
JUL 7
JUL 14
JUL 21
JUL 28
AUG 4
AUG 11
AUG 18
AUG 25
SEP 1
SEP 8
SEP 15
SEP 22
SEP 29
OCT 6
OCT 13
OCT 20
OCT 27
NOV 3
NOV 10
NOV 17
NOV 24
DEC 1
DEC 8
DEC 15
DEC 22
DEC 29
JAN 5
JAN 12
JAN 19
JAN 26
FEB 2
FEB 9
FEB 16
FEB 23
MAR 2
MAR 9
MAR 16
MAR 23
MAR 30
APR 6
APR 13
APR 20
APR 27
MAY 4
MAY 11
MAY 18
MAY 25
JUN 1
JUN 8
JUN 15
JUN 22
JUN 29
JUL 6
JUL 13
JUL 20
JUL 27
AUG 3
AUG 10
AUG 17
AUG 24
AUG 31
SEP 7
SEP 14
SEP 21
SEP 28
OCT 5
OCT 12
OCT 19
OCT 26
NOV 2
NOV 9
NOV 16
Aug 4, 2026
Earley AI Podcast - Episode 96: AI in Clinical Trials, the Vibe Coding Fallacy, and Bending Eroom's Law with Patrick Leung
16:46
P
14:09Patrick LeungGUEST
Well, just, just briefly, so the f- the four factors driving Eroom's law, number one is the fact that over time, inevitably the bar rises for new treatments because you have these...
P
14:21Patrick LeungGUEST
And so it's just requi-- it's harder and harder to come up with a treatment for a given indication, a given disease.
P
14:27Patrick LeungGUEST
Um, the second is there-- every once in a while there's a regulatory disaster.
P
14:30Patrick LeungGUEST
There's a thalidomide, there's a fentanyl, and then when that happens, the regulatory authorities become that much more stringent, and rightly so.

Seth EarleyHOST
Protocol optimization, so taking those protocols, making them sa-sa- uh, safer and cheaper and, and leaner by stripping out unnecessary activities and, uh, predicting the clinical outcomes, right? What, what, you know, where you have predictive models that are built on interlinked data sources that estimate the likelihood of technical success, safety issues, amendments, uh, you know, and, and, and, and, and the question is, so talk a little, a little bit more about those four areas where you're applying AI, and which one surprised you the most in terms of the challenges and the difficulty?
#26: How AI Accelerates Drug Development - Patrick Leung
51:07
P
24:40Patrick LeungGUEST
One thing I'll say is that there's a certain class of insights that we've had in the tool for some time that are straight up analytics.
P
24:47Patrick LeungGUEST
And so what we're doing is essentially looking at the whole structure of the trial and figuring out based on all the different patient visits and all the different patient arms and all this kind of thing, what's the overall cost going to be? What's the overall patient experience going to be? And this doesn't involve AI at all.
P
25:02Patrick LeungGUEST
This is just straight up sort of adding things up in a way that's in line with the design of the trial.
26 MINS LATER

Ian KrietzbergHOST
So I'm just curious where you land on this question of tools or synthetic life, digital God, you know, what are we building? What are we leading toward in your mind?
Saving Pharma Companies Billions with AI l Patrick Leung from Faro Health
R
12:25Raza HabibHOST
Yeah, it's, it's extremely staggering. And that's why I think I was a little bit surprised and, and fascinated to learn, like, that a big part of that cost actually is the logistics of it, rather than just the kind of pure science. And that, that suggests that really there is... If I'm reading that, I'm understanding correctly from you, Patrick, that means that, like, the leverage from using AI to automate these processes could really be very significant in bringing down that cost?
P
12:45Patrick LeungGUEST
Yeah. I mean, w- we definitely see a huge opportunity here because we... During the course of designing the clinical trial, we have this whole plan laid out. Over the course of, in, in many cases, like, many years, where you have a bunch of patients that are visiting a bunch of different sites and having a bunch of different assessments, you know, measurements, drug administrations, all these kind of things done to them. And so this is a really big, complex trial that involves people and resources and places and so on. And so the opportunity to actually optimize that logistically for various utility functions, whether it's sort of what we're seeing in- increasingly, people are worrying about carbon emissions, especially in Europe. Of course, cost is always an issue. Patient burden, like, how much time is the patient spending away from work? This is a really big factor in terms of your ability to re- recruit the patients, for them to stick with it, and so on. And so it's just, basically, there's a lot of things here that could be optimized. And I think that certainly having a structured repository that contains all this information is helpful. And I think that as time goes on, we'll think of more ways to apply AI to essentially optimize this process. So we talked before about generating documents. We're actually also doing the opposite. So out there in the world, there's tons and tons of trials that are already stored in public places like clinicaltrials.gov. You go to clinicaltrials.gov, you can download thousands and thousands of clinical trials. They're all PDFs. And you can sort of read about all these different trials that have been done. So when I came on board, I thought, like, "Well, what if we could actually parse those documents and extract structured information from them and enter them into our system and view the insights, and basically, in principle, be able to do a lot with the, a lot of interesting things with that data, such as compare them with trials that you're developing?" So a really common use case we have is a pharmaceutical company comes in and says, "Look, we're developing all these trials. We want to know what other trials out there are like this. Can you do that for us?" That's a lot of work. Like, that is a lot of manual labor to, to, to do that. Weeks of work, potentially, to find a cohort of benchmark trials and basically enter them in- in- into the system and do all the analytics. And so we're actually using AI to do that automatically, where we can go through the PDF, extract all the tables and images and text, vectorize all the text, and produce this very rich, very, very interconnected database, kind of like a graph database, that essentially we can use to introspect. And we can do queries like, "Well, given this trial that we're working on right now," maybe it's an oncology solid tumor trial, s- some kind of complicated trial, "give me some examples of trials that are like this, but have a lower patient burden, or that are, you know, like, have a shorter schedule, or give me some examples of assessments that I might be missing, or that I don't even necessarily need." So the- these are things that typically would take weeks of research, right? To, to just go through places like clinicaltrials.gov, download a whole bunch of interesting-looking PDFs, pore through them, realize that half of them are irrelevant, and come up with the necessary insights, right? And so we're, you know, in this process of developing technology where this can all be done automatically, and you can essentially sort of do this using a natural language interface within our tool. So I'm pretty excited about that because I think that it's really gonna open the door to much better clinical trials. Like, much more well-researched, well-thought-out, optimized trials that are better for patients, that are less likely to encounter delays with the FDA, and that will essentially cost less money to bring to market. And I think this is what's needed, right? Because, you know, I'm, I'm sure many of the people in this, this audience are aware that there's all these incredible innovations going on, right, in biotech. Like, our understanding of...... how the cellular membrane works and how genetic medicine can, can work. And all these new treatments are coming, and the trials just get more and more complex for these things. And so there's definitely a bottleneck when it comes to actually getting them through FDA approval, and that's really what we're here to, to solve.
10 MINS LATER
R
26:03Raza HabibHOST
The High Agency Podcast is sponsored by Humanloop. Humanloop is the LLM eval platform for enterprises. Teams at Gusto, Vanta, and McMillan use Humanloop to ship reliable AI products. If you're an engineering leader or product manager building an AI product, you can find out more at humanloop.com. Fantastic. So Patrick, that covers kind of use case one, which is, "Hey, take me, you know, take all this structured data and use it to create the unstructured documents that the FDA needs." Can you tell me a little bit about how that differs from kind of big use case number two of taking, you know, the unstructured documents and, and parsing them in the, in the reverse order?
P
26:37Patrick LeungGUEST
Yeah, I mean, they are almost completely different, uh, applications. Like, despite the fact that they're sort of mirror images of each other. Because now we're talking about taking, you know, like, PDFs are, as a standard, is very complex and sort of unpredictable. It's not like sort of HTML or, or JSON, where it's pretty straightforward what the, what the format is. So PDFs themselves are inherently quite difficult to actually parse and decompose. And so there's various libraries out there that we're using to do that, and they can pick out the images and the structured data in the form of tables and so on. But the hard part really is, okay, you've decomposed everything, all the text, the images, the tables, and so on. But of course these things refer to each other. So we can take the table, the, the SOA, the schedule of assessments, that contains a whole long list of all the different procedures, right, that are essentially gonna be performed during the course of this trial, usually dozens and dozens of procedures. But first of all, we need to normalize those, right? So, one of the things that Pharo has done as part of this tool is we've developed this whole library of standardized assessments. And so one of the first things we do, actually, is take the assessments that we find in the PDF SOA in, in the table, and match them up with our library so they can actually be normalized and compared. Like, that requires its own LLM. Like, we need to kind of do a classifier, essentially, that takes our, our BCL, Biomedical Concept Library, as being the truth set, and we try and do a matching. But then that's just the beginning, right? Th- so this is, this is standard sort of entity resolution stuff. But after that, we need to sort of figure out, okay, we have a basic schedule, you know, coming out of that table, but that is a compressed schedule. Like, oftentimes, in order to fit it inside the confines of an eight and a half by eleven page, a lot of, there's a lot of shorthand in there saying, oh, by the way, you know, there's a footnote. You know, footnote number one or footnote number J or whatever that says, "Oh, by the way, there's this certain other assessment that needs to be done every second day." Or, "Oh, by the way, like, after the, you know, the first three cycles, this now gets performed every week." And so there's all this kind of fine print that gets peppered, you know, throughout the, the, the clinical trial that we need to kind of sleuth. We need to kind of spelunk through the trial and find everything that refers to this assessment and then reconstruct a fuller, more detailed version of the schedule from there. Because we need that fuller, more detailed version of the schedule in order to calculate insights and be able to compare the trial with other trials and do all the good things we want to do with this data. So there's quite a complex pipeline we've built that essentially threads the needle between the table and all the references and the footnotes and all the references and other parts of the clinical pro- protocol document. And then, you know, of course, cross-references and matches with our own database and so on and so forth. So it's a very different type of problem. This is straight-up data extraction. It's not generative. It's extracting structured data from an unstructured source. And it's super exciting, you know, 'cause I, I feel like we're building a dataset that hasn't really been seen before. And we can do really interesting things with it. You know, first and foremost, just benchmarking. Like, being able to take a trial and sort of come up with a cohort of similar trials and just show stats and say, "Oh, you're in the bottom 25th percentile of patient burden," or, "You're in the top 25th percentile of time taken," or whatever the case may be, right? Like, it's, that's useful information that every sponsor is gonna wanna know about drugs under development. But then I, it gets more interesting than that because, again, we have LLMs, right? So LLMs are great at natural language processing and so on, and so we can go beyond these basic insights and say things like, you know the example I said before, right? "Given this trial, like, show me some similar ones that actually have a lower patient burden." Or, "Hey, you know, do trials like this, like, how's, how are things likely to go with the FDA anyway? Let's look at a whole bunch of trials somewhat similar to this and see how they fared, you know, as far as amendments. Like, what kind of amendments did they have, and like, h- what were those amendments anyway? Why did they run into trouble with the FDA? How can we avoid that? Hey, system, like, how can we avoid that?" And it's gonna come up with some, some suggestions. So this is taking clinical trial design to a whole other level of intelligence. And, you know, we're, we're automating research that, if it is being done at all, you know, takes, takes a lo- lot of extra effort. And, and these are highly paid, highly specialized individuals doing this work. So we're really powering them up to do their jobs a lot better and produce clinical trials that are, that are just way better for the patient and way... and be- better for everyone, essentially.