Outcomes Data from Clinical Applications... - Helen Colhoun

Helen Calhoun: So, thank you very much for inviting me to
give a talk here today. It has been a really interesting meeting and I must admit, I was
really struggling as to whether or not my talk was going to be on message or not. So,
I think it's approximately so anyway. Tim, I'd really like it if you'd get up and do
a little dance about two minutes before I finish. Okay? So, I'm a public health doctor
and epidemiologist and I probably know less about genetics than most of you in the room.
So today I wanted to talk to you a little bit about -- sorry, I can't get this working
to go forward.
Male Speaker: Left button --
Helen Calhoun: Okay, got it. Okay. So, three things that
struck me about yesterday were Professor Khoury saying we need to do the studies to provide
the evidence base for clinical utility. Then I think I'm paraphrasing here, so excuse me
if I'm putting words in your mouth, but I think it was Dr. Aronson that said if you're
going to put things in bin two you actually need to state clearly what you need to know
to get it out of bin two. And the third thing that I think Elaine said was that her priority
was to have, you know, better data to establish whether very rare variants are likely to be
causal. So that's a huge span of activities I think that needs to be considered.
Now the Ph.D. Foundation in the U.K., I think Tim and Paul were involved in this, recently
produced this really comprehensive report on the future with whole-genome sequence and
what it means to the NHS in the U.K. and one of the statements that they've made is that
the NHS presents a wonderful opportunity to implement whole-genome sequencing in a way
that's evidence-based, systemic and efficient. So I think what I was going to try to sort
of talk about today was how can NHS data really be used? What are the practicalities and some
examples about how such large scale data might be usable for answering some questions to
help close the translation loop? And I wanted to use MODY maturity-onset diabetes in the
young, which you all know as monogenic diabetes, as sort of a working example of that.
So I'll just briefly mention something about allotonic health care data available for research
in Scotland, managing our bioresources and how they link into data and then talk a bit
about MODY. It's very unresponsive. So the key points are that in Scotland we've had
a unique health care identifier for quite some time now available on all health care
records, so it follows the patient around with every health care encounter that they
have. Okay? And through to death records. And this unique identifier permits linkage
between many different available data sets. So for example, we can link together all hospital
admission records going back to 1981. We can link to maternal and child health care records.
We can link to psychiatric records. And in some circumstances, we can link to primary
care data, though the expert on primary care data is here in the room, so we can pick his
brain after, John Parkinson, who manages GPRD. Importantly, through a welcome trust funded
initiative, the Scottish Health Informatics program, a bunch of it's led by Professor
Andrew Morris who works in Dundee where I'm also based, have been working through a wide
range of issues, including importantly some of the governance issues and data safe haven
issues for how you actually collate and use those kinds of data.
To mention one or two bioresources that we have, because Rex said at the beginning, you
know, we need to really ask ourselves do we have enough actual established bioresources
and available data sets with good depth of clinical annotation. So here is one example.
Generation Scotland is a triumvirate really of three studies, but probably the most interesting
one is this one here, the Scottish Family Health Study, which is not a completed collection.
It has altogether about 24,000 patients in this study. They are actually sampled from
the general population. They have extensive and deep phenotyping done. It's a pedigree-based
structure, so it comprises about 7,000 families and importantly, we can link all of the data
to all of these other routine health care data sets that I mentioned. So it's an enormously
valuable resource for research and it's open for people to apply to use it.
Sorry, I'll just go back. Can I go back? Right click. That's what I'm doing, but it's not
working. Okay. Okay. So, there's -- as I say, a depth [spelled phonetically] of phenotypes
in here. I won't go through them all. So there are some study-specific instruments that are
used in Generation Scotland.
Now I'm going to turn to diabetes data. In Scotland we have about 250,000 people with
diabetes, about 25,000 of them have type 1. We have an electronic health care record that
is used throughout the country. It's used by most hospitals as its primary EHR, but
even where it's not there is a feed in to SCI-DC, as we call it, from the EHR that that
hospital is using. It also receives a nightly feed of key items from every primary care
physician in the country, bar 3, on quite a lot of data, including issued prescriptions.
So that's pretty much what I spend quite a lot of my time on is studies that are built
around this data set. We can link the data in this data set to other, say all of these
other record systems here, database systems here.
And importantly, we have built various buyer resources so the welcome trust-funded case
control by resource, for example, linked to these data in [unintelligible] has been pivotal
in a lot of the replication studies for the major finds of disease -- of genes for type
2 diabetes in the last few years. The type 1 bioresource is something I'm building at
the minute. We started collecting in February and we've bio-banked almost 3,000 patients
so far and our focus with this ultimately we want to get to 10,000 is really very much
around diabetic complications. [sighs] Sorry, I'm just not getting on with this at all.
[laughter]
Helen Calhoun: So, this is what happens when Tim tells you
to hurry up. Okay, so we have this bioresources, but you could just as easily substitute the
brave new world. What if you have next-generation sequence data on all these patients? And even,
what if patients get it themselves and want to upload it and append it to their clinical
record? Because we do have a patient facing aspect of the data set. So we might want to
think about that. So that's a bit about resources and, oh, God, two minutes. Okay.
So, MODY. Most of you know about MODY. Eighty percent of MODY is monogenic diabetes, but
it's usually clinically misdiagnosed as either type 1 or type 2. We've known about MODY and
we've known about the genes that underpin it for many, many, many years. Right? This
work by Andrew Hattersley's group, who's the world's leading expert on MODY, has shown
that we currently diagnose less than 20 percent of all MODY. Okay? I don't know how the figures
are in the U.S. or if anybody's looked. I know that EGAPP you have it on your list,
but you haven't done your review yet of it.
So, why? So it's a perfect example of an actionable but unactioned variant, and maybe we need
to take stop before we worry about all the stuff we're going to learn tomorrow about
how lousy we are at implementing what we know today. And so there's a whole bunch of issues
here in relation to this, which I won't go through, but the question is how do we do
studies that actually solve some of those issues? Now, first thing is there hasn't been
a cost benefit analysis, okay? But the different data elements that would be needed for that
still need to be generated. So at the minute some of the studies we're doing are using
our bioresource linked to our national data set and in conjunction with Andrew Hattersley
doing a study called the United Study where we're evaluating certain algorithms for prioritizing
who should get sequenced. Okay? And so that's an important thing and that's part of the
actionability as distinct from the clinical utility space. Okay?
So one of the real issues here is to bring down costs to improve that cost benefit ratio.
The question is, should you stratify first biomarkers and family history and clinical
features from the record? If so, how? Which ones are best? Which ones yield more? Which
ones are most cost effective? And interestingly, this is one area where you might want to think
about other biomarkers in your bio-bank being really useful for telling you something about
who needs to be sequenced. So fascinating aside here is that a GWAS of the plasma glycome
recently revealed HNF1 alpha, one of the main genes from MODY, to be a master regulator
of fucosylation, opening a whole field of using N-glycome branching assays as a diagnostic
test. So that's one example.
Another example where I mentioned randomized trials is exactly in this field of clinical
decision support. So one of the things we're trying to design at the minute, which we can
do, is not exactly randomized, but what we can do is we can implement in different parts
of the country a clinical decision support tool to prompt the potential screening for
MODY and we can compare then how well that's achieving an increased yield of cases in comparison
to the status quo. So that's something we're trying to design at the minute.
I've mentioned about the studies on looking at the different algorithms for how you might
approach something and then I think an interesting thing to go back to Elaine's question about
how to infer causality, we might want to consider what the future will look like if you've got
lots and lots and lots of sequenced data that you just happen to have on people, okay? And
you have lots of phenotypic information, how we best exploit that. And at the minute, I
think this is an idea that I picked up from my husband when I was talking to him about
coming here this week, because he's been doing some work on basically detecting, using GWAS
data, increased regions of IBD sharing. So if you have lots of GWAS data in your population,
you've already got some MODY cases that have been sequenced. You could look at the IBD
sharing and actually say hang on, some of these patients who appear to be type 1 are
actually have access sharing with some, I know, MODY cases. Is this a route to detach?
And so you could do a whole sweep of your population to evaluate people. Anyway, it's
just a thought, but I think it's worth thinking through.
Summary and conclusions. We need to harness the power of EHRs, link to bioresources to
complete the translation loop. We can do the clinical validity and utility studies, but
that needs money, and we can also think about the more actionability questions, including
actually doing randomized comparisons of approaches to actionability, but it needs demonstration
projects and systematic effort and I would say with some careful consideration about
the feedback of WGAS [sic] data at the minute in these situations. And finally, any effective
reporting back need to be formally evaluated so as to feed back into clinical utility.
I'll stop there. Sorry about the slide dancing.
[applause]
Male Speaker: The -- I'm interested in the consent that
you operate on. Is all the data anonymized or is there -- do you have access to identifiable
data and what consent were these collected under?
Helen Calhoun: Okay, so we have two different systems that
operate, okay? We have one system if you just want to do what I call dry data analysis where
you're linking records together and they're completely anonymized, deidentified, et cetera.
For that what we do is we have a system whereby we have a privacy guardian, as well as ethics
committees, so we have to get a privacy guardian approval from everybody, but also increasingly
now what we've set out is a blueprint for a maximal secure way of utilizing those data.
So even when they're deidentified, our new system is going to require passports for validated
researchers. It's going to require the data to reside within data safe havens and so forth,
so that's that level.
Then in terms of bioresources, they are all by definition individually consented. So the
patient comes in and we consent them for collection into the bioresource, but we also consent
them for retrospective and prospective linkage to their clinical record, okay? And then a
third form of consenting that we do with patients is ad hoc, as patients come into the clinic,
we consent them also into our research register, which allows us to provide them directly with
information about studies we're trying to recruit from without necessarily having to
go through their primary care physician again.