Going about things in the right way

I got involved with Medisoft through the retina side – just talking to Rob and seeing the fantastic EMR analysis work they were doing on cataract surgery and what they’d been able to show in terms of the impact of the introduction of anti-VEGF agents. We could see that it would be a really important area to look at – these are expensive, intensive therapies, and we really need to know that our patients in the real world are doing as well as those in the clinical trials. I managed to convince a pharma company to cough up a trivial amount of money to pay for the data collection, and we did the rest ourselves. And within two months of thinking of the idea, we wrote to 16 Trusts, got replies from 14 within a month, and a month after that we had data from Moorfields. Imagine that – just two months after having the idea we had outcome data on over 100,000 injections, representing 300,000 patient visits and about 2.8 million data points!

Our first attempt was in 2011; at that time, our statisticians were not used to handling big datasets, so I actually did it myself on my laptop. I am probably very fortunate in that, although I am not a proper “techie,” I was lucky to have been given a computer – a BBC Micro – as a kid, and I learned to do basic coding. But then I went to medical school. I have always been okay with numbers, but medical school knocks it all out of you! When I started doing my research, I got back into analysis. The Medisoft EMR data is very structured, but it is very difficult to deal with big datasets and link them together in something like Excel, so I wrote some basic scripts in an SPSS software package, which took me a few months to do, during my holidays in Australia. Anyway, we got some really good data out – and the first two papers were published in Ophthalmology (4)(5).

Everyone now realizes that you need to treat early and it is all about visual acuity state, not gained visual acuity. If you have poor vision, let’s say in AMD, and then treat it, the average patient will gain a lot of letters; but if you have good vision, the average patient loses vision. And the way that most people audit outcomes or the way that the data is presented in all the trials on ANCHOR and MARINA is visual acuity gain from baseline – but that isn’t what’s important to the patient. What is important to the patient is visual acuity state: not whether they have gained two letters, but whether they can still drive. It is all about getting the patient early. And it has huge implications, because it really influences clinical trial design – non-inferiority trials, superiority trials, because you can almost game the population you want to enter. So there were huge implications even in that first paper – it’s not just about asking, ‘how are we doing?’

And we realized that, at least in the UK, we weren’t giving as many injections as we thought we were (4). The vogue was not to over-treat because we were worried about the injections and the side effects, so there was a natural reluctance. The outcomes were not disastrous, but nothing like ANCHOR and MARINA – and with very few injections. So we realized we probably needed to alter the way we do things. In the second paper on second eye involvement (5) – I think Javier was the first author of that – we realized that the risk of the second eye succumbing was actually much greater than we thought. Once you start treatment in one eye, if you have reasonably good vision in the fellow eye, you have a 50 percent risk of developing wet AMD within three years. Again, this has huge implications in terms of the extent of treatment required, and what happens when both eyes are involved.

I did the analysis of the first two myself – but it was pretty tiring to do the day job as well. I realized that it would be nice to get much better coders involved and I was very fortunate to meet some extremely good data guys, who also understood the eye. There was Dave Crabb’s group at City University in London, and a fellow, Aaron Lee – a rare individual who not only codes brilliantly but is also a retina specialist. So we then developed a group with me, my wife [Cathy Egan], Dave Crabb, and Aaron – and then it really took off; my ability to code was no longer the limiting factor. Once we had brilliant people on board, we just started free-thinking about what we could do, and not just limiting ourselves to conventional outcomes. For example, we started looking at novel health economic analysis, such as how in the UK at the moment you still can’t treat eyes with vision better than 6/12 (20/40).

With health economics analysis, you need to establish that what you are doing is incrementally more cost-effective than the standard of care. But the standard of care is unknown because in the trials there is no control arm better than 6/12. It was Rob who bounced the idea around and realized that the data probably existed in Medisoft. There were a few areas in the UK where they were funding vision better than 6/12, so we mined that fellow-eye data, and their OCT data, and when we had a leap on the OCT for two successive visits, we followed it through until they got an injection. That was the natural history arm; then there were centers that had the injection, and that was the real arm. So we then did a health economics analysis, and showed it was highly effective even with an ICER analysis; there is a whole level of complexity above that. It has been studied in the current NICE AMD review, and we hope it will change things.

We have done a similar study now with AREDS, which was published recently (6). The UK doesn’t offer AREDS vitamin supplementation. The AREDS formulation is known to be effective for preventing wet AMD in people who’ve got the wet form in one eye and the dry form in the other, and we were very lucky to do a hybrid health economic model. Emily Chew of the NEI gave us full access to the original AREDS dataset; we merged that with real-life outcomes and got a NICE-level health economic model. It turned out that AREDS aren’t just “cost-effective,” if you’ve got wet in one eye and dry in the other – they are what’s called “dominantly cost-effective.” In other words, if you pay for it, it still saves the NHS money – to the tune of £130 million a year per cohort you enter. It wouldn’t be possible to show that without our combination of data. I believe the approach is going to be transformative, not only for outcomes – by ensuring that we are modifying our behavior in the UK – but also for health economics.

We also have the potential to “free think.” I had a fellow operating on my list, a brilliant surgeon, and he had a complication. We looked up the EMR records and found that the patient had received about 22 anti-VEGF injections. A few weeks later, the same thing happened. We looked up the notes and found that the patient had received 30 injections! I thought, we’re injecting the eye, we’re not seeing obvious trauma, but this may be because you are injecting and somehow weakening the lens. As I said, this is a rare event, and it has been very difficult to prove the association – even when it’s the two most common procedures in ophthalmology – cataracts and injections. Within 48 hours of asking the question, Aaron analyzed the whole dataset (because he had access to the cataract data and rupture rates) and realized that actually there is an increased risk per intravitreal injection: when you have more than 10 injections, your risk of complications or posterior capsule ruptures is the same as a junior surgeon or a resident surgeon – it has that much of an impact! Such knowledge is important as it changes how you advise your patients. We realized that we can actually do more than simple outcomes. We are now at the next stage: looking at different procedures within the eye, and how they interact – and trying to understand how that affects the rest of the body.

Now, the challenge is merging other public EMR datasets, which is ethically complex; however, we are now linked to the Farr Institute, one of the four government-funded Health Informatics Institutes, to link eye data to cardiovascular data to GI data. We can start looking at uveitis (we did an editorial on big data in uveitis in Ophthalmology not so long ago [7]), about how it interacts with the body. What drives machine learning and trains a deep-learning algorithm (or any machine learning algorithm) is lots of data. Now, we are using our data to drive the next generation of machine learning, and not just in terms of analysis of images. We’re working with a number of groups: we are linked to the University of Washington, where Aaron is now, and we are part of the last ever big European grant. We are also doing more complex things; we have a very large pharma grant of about £2,000,000 to link imaging to the EMR data in search of the Holy Grail: can we predict what is going to happen to a patient, and select them out? In other words, if we’ve got a predicted poor responder on the basis of various factors, can we switch them onto other clinical studies or other therapies? We have 43 million OCT slices from the service at Moorfields now linked to the EMR data, which is going to drive all of that effort. It’s our next big challenge.

And all of this was inspired by Rob – none of it would have happened without him. He has also inspired other important research in this area. On that basis, I thought it would be a good idea to do a Special Interest Group at ARVO a few years ago; I invited Mark Gillies – because he was doing some FRB, and we also talked about cataracts – and I invited Dave Crabb to talk about glaucoma, and Aaron to talk about data visualization, and the group really took off. Then, the American IRIS registry group came on board, and kind of took responsibility for it last year, and invited me to be on an all-day big data course. Many pharma companies were present, and now they’re all interested in big data research.

This whole evolution of data-driven research was triggered by Rob’s work. Without him, the tipping point would probably have been many years later.