Thinking outside the cylinder: on the use of clinical trial registries in evidence synthesis communities

Clinical trials take a long time to be published, if they are at all. And when they are published, most of them are either missing critical information or have changed the way they describe the outcomes to suit the results they found (or wanted). Despite these problems, the vast majority of the new methods and technologies that we build in biomedical informatics to improve evidence synthesis remain focused on published articles and bibliographic databases.

Simply: no matter how accurately we can convince a machine to screen the abstracts of published articles for us, we are still bound by what doesn’t appear in those articles.

We might assume that published articles are the best source of clinical evidence we have. But there is an obvious alternative. Clinical trial registries are mature systems that have been around for more than a decade, their use is mandated by law and policy for many countries and organisations, and they tell us more completely and more quickly what research should be available. With new requirements for results reporting coming into effect, more and more trials have structured summary results available (last time I checked, it was 7.8 million participants from more than 20,000 trials, and that makes up more than 20% of all completed trials).

The reality is that not all trials are registered, not all registered trials are linked to their results, and not all results are posted in clinical trial registries. And having some results available in a database doesn’t immediately solve the problems of publication bias, outcome reporting bias, spin, and failure to use results to help design replication studies.

In a systematic review of the processes used to link clinical trial registries to published articles, we found that the proportions of trials that had registrations was about the same as the proportion of registrations that had publications (when they were checked properly, not the incorrect number of 8.7 million patients you might have heard about). Depending on whether you are an optimist or a pessimist, you can say that what is available in clinical trial registries is just as good/bad as what is available in bibliographic databases.

Beyond that, the semi-structured results that are available in are growing rapidly (by volume and proportion). The results data (a) help us to avoid some of the biases that appear in published research; (b) can appear earlier; (c) can be used to reconcile published results; and (d) as it turns out, make it much easier to convince a machine to screen for trial registrations that meet a certain set of inclusion criteria.

I suspect that the assumption that clinical trial registries are less useful than the published literature is a big part of the reason why nearly all of the machine learning and other natural language processing research in the area is still stuck on published articles. But that is a bad assumption.

Back in 2012, we wrote in Science Translational Medicine about how to build an open community of researchers to build and grow a repository of structured clinical trial results data. The following is based on that list but with nearly 6 years of hindsight:

  • Make it accessible: And not just in the sense of being open, but by providing tools to make it easier to access results data; tools that support data extraction, searching, screening, synthesis. We already have lots of tools in this space that were developed to work with bibliographic databases, and many could easily be modified to work with results data from clinical trial registries.
  • Make all tools available to everyone: The reason why open source software communities work so well is that by sharing, people build tools on top of other tools on top of methods. It was an idea of sharing that was borne of necessity back when computers were scarce, slow, and people had to make the most of their time with them. Tools for searching, screening, cleaning, extracting, and synthesising should be made available to everyone via simple user interfaces.
  • Let people clean and manage data together: Have a self-correcting mechanism that allows people to update and fix problems with the metadata representing trials and links to articles and systematic reviews, even if the trial is not their own. And share those links, because there’s nothing worse than duplicating data cleaning efforts. If the Wikipedia/Reddit models don’t work, there are plenty of others.
  • Help people allocate limited resources: If we really want to reduce the amount of time it takes to identify unsafe or ineffective interventions, we need to support the methods and tools that help the whole community identify the questions that are most in need of answers, and decide together how best to answer them, rather than competing to chase bad incentives like money and career progression. Methods for spotting questions with answers that may be out of date should become software tools that anyone can use.
  • Make it ethical and transparent: There are situations where data should not be shared, especially when we start thinking about including individual participant data in addition to summary data. There are also risks that people may use the available data to tell stories that are simply not true. Risks related to ethics, privacy, and biases need to be addressed and tools need to be validated carefully to help people avoid mistakes whenever possible.

We are already starting to do some of this work in my team. But there are so many opportunities for people in biomedical informatics to think beyond the bibliographic databases, and to develop new methods that will transform the way we do evidence synthesis. My suggestion: start with the dot points above. Take a risk. Build the things that will diminish the bad incentives and support the good incentives.

Differences in exposure to negative news media are associated with lower levels of HPV vaccine coverage

Over the weekend, our new article in Vaccine was published. It describes how we found links between human papillomavirus (HPV) vaccine coverage in the United States and information exposure measures derived from Twitter data.

Our research demonstrates—for the first time—that locations with Twitter users who saw more negative news media had lower levels of HPV vaccine coverage. What we are talking about here is the informational equivalent of: you are what you eat.

There are two nuanced things that I think make the results especially compelling. First, they show that Twitter data does a better job of explaining differences in coverage than socioeconomic indicators related to how easy it is to access HPV vaccines. Second, that the correlations are strongest for initiation (getting your first dose) than for completion (getting your third dose). If we go ahead and assume that information exposure captures something about acceptance, and that socioeconomic differences (insurance, education, poverty, etc.) signal differences in access, then the results really are suggesting that acceptance probably matters, and that we may have a reasonable way to measure it through massive scale social media surveillance.


Figure: Correlations between models that use Twitter-derived information exposure data and census-derived socioeconomic data to predict state-level HPV vaccine coverage in the United States (elastic net regression, cross-validation; check the paper for more details).

It took us quite a long time to get this far. We have been collecting tweets about HPV vaccines since 2013. And not just the (well over) 250K tweets; we also collected information about users who were following the people who posted those tweets. Well over 100 million unique users. We then looked at the user profiles of those people to see if we could work out where they came from. Overall, we located 273.8 million potential exposures to HPV vaccine tweets for 34 million Twitter users in the United States.


Figure:  The volume of information exposure counts for Twitter users identified at the county level in the United States, by percentile, up to 19.7 million in New York County, NY.

Each tweet was assigned to one of 31 topics, and then we determined state-level differences in exposure to those topics. For example, Rhode Island was much more often exposed to tweets about legal changes in Rhode Island. Users in New York State had higher proportional exposure to tweets related to a series of articles published about the representation of HPV vaccines in the Toronto Star. States like West Virginia and Arkansas had a greater proportion of exposures to tweets about a controversial story that aired on a television show hosted by Katie Couric. A higher proportion of the exposures in Kentucky, Utah, and Texas were related to politics/industry rhetoric related to vaccine policy in Texas.


Figure: The five topics with the strongest negative correlations with HPV vaccine coverage at the state level in the United States (2015 National Immunization Survey). Mainstream news media features heavily among the topics with the strongest correlations.

The tl;dr—by characterising the information that a sample of Twitter users from each state were exposed to during a two year period, we were able to predict which states would have higher or lower rates of HPV vaccine coverage in the same period of time.

An extra note on computational epidemiology research

We hear about quite a number of studies that use social media data (Facebook, Twitter, Instagram, Reddit, Foursquare, etc.) to try and answer health-related research questions. Despite the volume, only a tiny handful of them have demonstrated a clear link between the kinds of data we can gather from social media and real world health outcomes at the population level.

The one that was most closely related to what we did with these tweets recently was published in Psychological Science and predicted heart disease mortality at the county level in the United States based on language used on Twitter. Eichstaedt and colleagues found that topics that were related to hostility, agression, boredom, and fatigue were positively correlated with heart disease mortality, and topics related to skilled occupations, positive experiences, and optimism were negatively correlated with heart disease mortality. They too found that socioeconomic predictors were less powerful in explaining differences in mortality than Twitter.

Other population-level research linking information posted on social media with real world health outcomes tends to fit in the scope of predicting influenza rates by location. I won’t go through all of these here, but these are probably the best represented in the research literature.

There are a number of other studies that use social media to predict health outcomes (or diagnoses) at the individual level, and these tend to consider who users communicate with and the language they use in their posts to predict things like major depressive disorders, alcohol use, suicidal ideation, and other psychological conditions.

So what does it all mean?

First, the research presented in the our new article is the computational epidemiology equivalent of basic research.

The research cannot tell us whether news and social media influences the vaccine decision-making of a population, or if it reflects people’s preference for reading information sources that they already agree with. It is probably a bit of both.

Second, the way we might operationalise the methods we have started to develop here is important. It is fine to know which topics appear to resonate with certain communities by doing location inference and demographic inference over user populations in social media [and just wait until you see what we are doing with location inference on Twitter and user modelling for conspiracy beliefs in Reddit]. But there is still a lot of hard work to be done to turn the information into actionable outcomes; informing the design of targeted media interventions and then tracking their effectiveness in situ.

For vaccination, we are still grappling with the evidence about how access and acceptance influence decision-making at the individual and population levels. For social influence, we still haven’t worked out the best ways to measure the contributions of homophily, contagion, and external factors to changing behaviours or opinions. We still don’t know whether social media data can be leveraged to better understand how media interventions influence other kinds of health behaviours. We still don’t know how much actual evidence about safety and efficacy makes its way into the information diets of various populations, and we don’t know whether populations are capable of distinguishing between high quality clinical evidence and the academic equivalent of junk advertising.

What we do know is that if we want to promote cost-effective healthcare, we need to continue to push for more research into the complicated cross-disciplinary world of preventive medicine—that weird mix of evidence-based medicine, epidemiology, social psychology, public health, journalism, and data science.

Why the inequalities in our information diets matter (Part 2 of 2)


Part II: A new NHMRC Project for measuring the impact of social and news media on health behaviours

As promised following Part I – and now that I am back from the burnt orange colours of the United States to the purple jacarandas of Sydney – another update. But first, a quotation from one of the books in a series that inspired most of the work I do. It starts with the Emperor of the Galaxy talking to a research academic:

‘I am given to understand that you believe it is possible to predict the future.’

Seldon suddenly felt weary. There was going to be this misinterpretation constantly. Perhaps he should not have presented his paper. He said, “Not quite, actually. What I have done is much more limited than that…”

We were lucky enough to have been awarded a new 3-year NHMRC Project. With co-investigator Julie Leask and a team of excellent associate investigators from Sydney and Boston, I will be developing a pipeline of interconnected methods capable of linking and tracking exposure to news media at unprecedented scales; estimating differences in information diets by geography and demographics; correcting for biases in each of the data streams we will mine; and connecting all of that information to real health outcomes. In the first instance, those outcomes will be related to vaccines…but we will be rolling out these methods into a range of other health outcomes as soon as we can.

Our major aims for the project are the following:

  • To determine whether measures of exposure to news and social media can be used to explain geographical differences in health behaviours (especially vaccines).
  • To determine the proportion of our information diets made up of evidence-based information, and characterise the quality of evidence in the news and social media that makes up our information diets.

If you are looking for an opportunity as a postdoctoral research fellow, and this kind of research sounds like fun to you, please get in touch. You will need to be very comfortable with developing machine learning methods for big messy datasets (GPUs and neural networks are a bonus), and have published innovative methods and released your code open source. And if you are considering a PhD, want a generous scholarship, and you can design a project that might align with these ideas (no matter whether your disciplinary background is computer science, medicine, epidemiology, statistics, psychology, sociology, or just about anything else), send me an email with your bio and ideas.

Why the inequalities in our information diets matter (Part 1 of 2)

Part I: Visiting the United States in November 2016

As I write this, I am on a train travelling from Boston to New York City, at a time when people in the United States are still coming to terms with what they thought they knew about their country. The trip has been rushed because I need to be back in Sydney to catch up on work, which meant I didn’t have time to catch up with all the excellent people on the east coast. Here is a picture of approximately where I am right now.

Between Boston and New York City, about an hour before New Haven.

Without the benefit of hindsight, it feels like the disintegration of knowledge continues to accelerate into the post-truth era; that societies around the world are struggling to hold onto their values across a populous that outsources its opinions and beliefs to some nebulous idea of whoever is both influential and accessible. While information overload is not at all a new idea, we do not yet entirely understand how information overload may actually restrict rather than open up what we hear, what we click on, what we soak up into our worldview; and how that can reinforce our echo chambers to create polarisation and conflict.

While it is not something we usually like to admit, the information to which we are exposed – our information diets – are an excellent predictor of our attitudes and opinions. This is not necessarily because our information diets causally affect what we believe and how we act, but because we surround ourselves with people who agree with us and rarely seek out opinions that are contrary to our own. And even when different opinions are put right there in front of us, we disregard them.

For researchers who like to observe the world while pretending not to be part of it, right now looks less like a post-truth era and more like a golden era for understanding the impact of news media and social structure on human decision-making and behaviour.

There has never been a better time to take advantage of massive streams of data about information search and exposure to measure their impact on opinion, attitudes, decisions, and behaviours. In my team, we already use social connections on Twitter to train machine learning classifiers that can predict whether you are likely to have negative opinions about vaccines. We are now close to finalising work showing that a population’s (relevant) information diet can explain the variance in vaccine coverage of individual states in the United States.

img_2087More of Boston walking from Back Bay to Boston Children’s Hospital.

Part II: Some news
While I am still looking at the bright orange colours of the United States, and before I head back to the purple jacarandas in Australia, I will stop here. But I will share some related news soon.


Five tips for controlling the evidence base of your clinical intervention

You might remember me from such articles as Even Systematic Reviews are Pretty Easy to Manipulate” and I Can Predict the Conclusion of Your Review Even Without Reading It“. If you have been around for a while you may even remember me from Industry-based Researchers get More Love Because They are Better Connected“. 

In a web browser near you, see the newest installment: More Money from Industry, More Favourable Reviews

With colleagues from here in Sydney and over in Boston, I recently published the latest in a string of related research on financial competing interests in neuraminidase inhibitor research. This is probably the last paper we will do on this topic for a while (though we do have two more manuscripts related to financial competing interests on the way soon). In this one we looked at non-systematic reviews of neuraminidase inhibitor evidence and compared the number of non-systematic reviews and the proportions of favourable conclusions between authors who had financial competing interests and authors who did not. You will not be at all surprised to learn that authors who had relevant financial competing interests and wrote non-systematic reviews about the topic ended up writing more of them, were more likely to conclude favourably in the reviews, and also wrote more of other kinds of papers in the same area.

So after looking in way too much detail at the creation and translation of evidence in this area, I thought it would be good timing to write down a few tips on how anyone with deep pockets can control an evidence base and get away with it (for a while). So here are some hints on the easiest and fastest ways to control the research consensus for a clinical intervention, even when it isn’t as effective or safe as it should be.

Step 1. Design studies that will produce the conclusions you want.

Step 2. When publishing trial reports, leave out the outcomes that don’t look good; or just don’t publish them.

Step 3. When publishing reviews, just select whatever evidence suits the conclusion you like best and ignore everything else.

Step 4. If the data fail to illustrate the picture you want, you can report them as is and just write down a different conclusion anyway.

Step 5. Use the credibility of reputable and prolific academic researchers by paying them to run trials, write reviews, and talk to the media.

Step 6. Profit.

Two important caveats: I am not claiming that any of these things have been done deliberately for neuraminidase inhibitors or any of the interventions described above – I am describing these processes in general based on multiple sources, and in a flippant way. Of course it might have happened for some or many clinical interventions in the past but that is not what we claim here or anywhere else. Secondly, I am not anti-industry, I am anti-waste and anti-harm.

And everyone should share the blame. There are researchers from inside industry, outside industry with industry funding, and completely divorced from all industry activity who have each been responsible for the kinds of waste and harm we read about after the damage has been done.

No matter what kind of intervention you work on, poorly-designed or redundant studies are a waste of money, time, and can put participants at risk for no reason. Failing to completely publish the results of trials is just as bad, and producing piles of rubbish reviews that selectively cite whatever evidence helps prove your preconceived version of the truth is about as bad as trying to convince people that a caffeine colon cleanse cures cancer.

When I find time, I will continue to add in related links to specific papers (for now, mostly just those from my team and my collaborators) for each of these areas. There are hundreds of other relevant articles that have been written by lots of other smart people but for now I am just listing a selection of my own as well as some of my favourite examples for each category.

So you’ve found a competing interest disclosure. Now what?

Published research varies across a spectrum that at one end is simply marketing masquerading as genuine inquiry. Actors in lab coats. To counter this problem, every time research is published in a journal, the authors are expected to declare anything that might have affected their impartiality in that work. Unfortunately, we very rarely do anything with those disclosures. It is as if by disclosing a potential competing interest, any effects on the research are supposed to either magically disappear or readers will somehow be able to magically account for their influence on the conclusions.

Let’s say you are reading a news story about a clinical trial that shows multivitamins will help your children grow up to be smarter, taller, healthier, and stronger. Seems to good to be true? You ask: “Is there a chance that the research has been distorted to make it look better than it really is?” so you try to find a link to the actual article so that you can check to see who the researchers are… and evaluate the quality of the research to determine its validity in context.

It’s actually much harder to do than it sounds, because for a substantial proportion of published articles, authors have failed to disclose things that would be defined as a potential competing interest by any standard definition. And in most cases, the competing interest disclosures are hidden behind paywalls, so you won’t be able to check the disclosures unless you have a subscription (or pay to “rent” the article, or use Sci-Hub to access it for free).

Then you ask: “What should I actually do if I encounter a competing interest disclosure?” At the moment, you have one of the following options: (a) you could either ignore the disclosure and take the research at face value; (b) you can throw the research out and ignore it because the research findings may be compromised; or (c) you could apply the best tools we have available for measuring the risk of bias even though we know they won’t always catch the nuanced ways in which research designs and reporting can be distorted.

Or more simply:  ¯\_(ツ)_/¯

So while we know that competing interests cause the kinds of biases that can lead to widespread harm, they also introduce substantial uncertainty into biomedical research because we simply don’t know if we can safely use research findings to inform our decision-making regardless of whether the authors have disclosed their potential competing interests or not.

But I think a solution to the problem is on its way. The first and most important step in that solution is to bring competing interests disclosures out into the open in ways that we can actually use. We need them to be made public, structured in a proper taxonomy instead of a bunch of free text descriptions, and accessible – in ways that both humans and machines can interpret and make use of.

That is why we think a comprehensive public registry of competing interest disclosures for all researchers is a good idea. People have been working on similar ideas for a while, and we have put many of these things in the review we published in the new journal, Research Integrity and Peer Review.

  • If you are interested, be sure to check out the IOM report (check the review), and read between the lines to try and understand why there might have been some disagreements about the best ways to deal with competing interests. This work eventually led to Convey, which has some similar goals but is not necessarily aimed at being comprehensive, and seems to be progressing nicely towards a different kind of goal from the updates on the webpage.
  • One of the things we didn’t include in the review because I hadn’t seen it until too late, is this browser plugin, which uses information from PMC to display funding and competing interests statements alongside the abstracts in PubMed. You could always click on the PMC link and scroll down to try and find them, but this is a neat idea.
  • It turns out that the idea for creating a standard list of competing interests and maintaining them in a public space was proposed as early as 2007, by Gordon Rubenfeld, in a letter to The Lancet. Maybe it is finally the right time to do this properly.
  • If you are a researcher and you like what you see in the first issue of the Research Integrity and Peer Review journal, then please consider publishing your research there. Besides the very interesting and high-profile editorial board, you might even have your manuscript handled by me as an associate editor.

Of course there are more things that will need to be done once we can manage for a much more comprehensive, consistent, and accessible way to disclose competing interests in research, but those come down to treating competing interests just like any other confounding variable. We are currently working on both methods and policies that will help us populate the registry in a longitudinal fashion (i.e. a timeline for every researcher who has published anything in biomedical research), and keep it up to date. We are also working on ways to take existing free text disclosures and classify them according to how much of an influence they have had over research findings in the past, and on a scale that has been virtually impossible until very recently.

Also, check out a more precise description of the idea in this related opinion piece published in Nature today. As usual, I will add to this post and link to any responses, media, comments, interesting tweets, etc. below as I spot them online.