How to read a VDU graph...

I'm a pretty simple guy. So the stuff that I put onto Virology Down Under's (VDU) blog is usually something I think can be understood by you - my yard stick is that if I can understand it, then I think you can. Sometimes it can get pretty technical though and with things always done in a rush, I don't stop and explain as much as I could. Which is why I value feedback. And I've had some good stuff from @DeclanButlerNat, @JorgeCastillaE and @Moro_Cedric this week. 

Different levels of experience read this blog and my posts on Twitter, so sometimes I direct my graphs towards them. But I do understand that we scientists can be easily carried away by our interests and forget that we're quite used to interpreting our own presentation styles in a certain and speedy way. We've had lots of experience doing it that way. I can change a tyre (as I was reminded a couple of nights ago, at midnight) but I couldn't fix my engine.

At the heart of reading a graph is this fact: you have to look at the axes to understand what the lines or bars or areas mean. Once you know the style, you can understand it at a glance - but first time, examine it with care. If it's one of mine, feel free to ask me what I'm trying to show if it is not immediately obvious. I very well may have failed to make it clear.

So this is a little overview of how to read some of the graphs which I use to communicate what I consider to be otherwise yawn-inducing tables of numbers about viral infection and disease numbers.

A picture is worth a thousand words..

This is a good thing because with my lack of typing skills, if I had to type 1,000 word all the time, that would be at least 200 typos. Graphs plot those tabular numbers in a more colourful and visual way. Once you know how to read a graph, they can become powerful and quick ways to get a quick update on the state of play. On VDU the game seems to be about outbreak data. That's just the way things have evolved for me since I first blogged on 28-March 2013. This includes graphing the number of people with disease (cases), changes in the number of cases, numbers that are suspected versus the number that are actually laboratory confirmed (my currency), dates of onset illness (favoured piece of data and the hardest to come by publicly), the numbers who die, the proportion (%) of all cases/detections who die, dates when disease was reported, sex, age and all of that can be plotted on graphs by day, week, month or year.

Interpreting a basic graph on VDU...

The graph below (Graph 1) comes from following Middle East respiratory syndrome (MERS) public data. It shows the key parts of the structure of the graphs - the axes (the horizontal and vertical lines that are the key to reading the plotted numbers) and the axes.

• A basic graph has a bottom horizontal line called the x-axis and it has a vertical line on the side called the y-axis. These are used to tell you what the numbers plotted on the graph mean; they are a key to the placement of each point on a graph, according to at least 2 different values.
• Each point on a graph represents a coordinate. Its made up of an x-axis values (abscissa) and a y-axis value (ordinate). For example we plot 50 cases reported on Thursday or 50 on the y-axis and Thursday one the x-axis (x,y)
• The points that we plot as pairs of x and y data can be joined up and shown as a line (the area underneath the line can also be coloured in which looks like a mountain that may have peaks and troughs) or they can be plotted as bars. There are other ways too - but I keep it simple. Joining up these dots is not always accurate - we may have no idea what is really happening to the numbers between any 2 points, in that case a bar graph may be more realistic as it shows the numbers at a distinct point in time. Sometimes bar graphs don't work from a formatting perspective (eg bars get so skinny you can't see them). Other times, joining the dots reveals the trends (the general direction that events are heading even if we don't know the values). Trends are useful in infectious disease as they show what has happened and what the latest data mean in the context of what has come before - so not too unrealistic. Some of this is about being accurate while not being too overly obsessive.

The particular example graph I've included below (Graph 1)  is a little trickier than some because it has 2 y-axes (vertical lines) - a primary (left-hand side) and a secondary (right-hand side). Some of the numbers are plotted against the primary y-axis (left vertical line) and some against the secondary y-axis (the right hand vertical line). This lets me "double-dip" on shared x-axis numbers, in this case, dates. I'm graphing the course of 2 different things (number of actual cases by day of illness onset) and the number of reported detected by date. These are 2 different things that have dates in common. 

This graph lets us compare, using the same x-axis, what the MERS case numbers look like when they are plotted by the day the people were reported to have become ill compared to the date of public reporting of the cases. There are differences that become more clear when you can run the 2 lines on the same graph, that may be a bit harder to see when they are plotted on 2 separate graphs. This graph highlights that when cases become ill and when they are reported are different things. It also shows that there were a bunch of cases (113) reported in 1 day that have never been given dates of illness onset (or hospitalization or the date they were each reported to the Ministry of Health). It also makes use of the 2 y-axes to have different scales. The primary or left-hand y-axis goes up to 35 while the secondary or right-hand y-axis maxes out at 120. If the same axis values were used, the illness onset cases would mostly be hard to see.

Graph 1. The basics of a graph.

What about cumulative graphs? What are they and how do I interpret those?

The next graph is made to show cases piling up over time (Graph 2). This is the graph that sparked this blog. It plots numbers as a line graph but instead of showing the value at that timepoint (day, week, month, year), it adds the new number to sum of all the previous numbers. It is plotting a cumulative tally, so it will always be a hill with an upwards (left-to-right, bottom to top) slope except when there are no new cases to add, when the curve becomes parallel to the x-axis - a flat line. How steep that line is can tells us how rapidly cases are piling up. That can also be fudged if you present the chart with a very short or long x-axis.

• In the case of the Zaire ebolavirus outbreak in West Africa, we have the unusual ability to compare numbers from multiple countries at the same time, and use the same x-axis. Here, we show the date when the World Health Organization's Disease Outbreak News update was released. Sadly for us graph addicts, this doesn't include any illness onset dates, but the WHO do have those data and plot it themselves here (1).
• A steep slope indicates a rapid rise in cases and this results from a lot of new cases being added in a short period of time.
• A near flat or horizontal slope to the line shows that there are not many new cases being added. 
• In this graph we also show multiple lines plotted using the primary (left) x-axis to present how much and at what rate the total suspect, probable and laboratory confirmed case numbers are piling up (pink) as well as how the deaths from among that number are changing (blue line) and how many of the cases are being laboratory confirmed (green line) as due to the virus suspected of being the cause. This last one is important as it gives a glimpse of how the laboratory network is coping, perhaps how specimen access is going and how much faith to put in the other two totals. Why are we worried about the result totals? Because many other things can look like Ebola virus disease (EVD) early on, and even later in the disease course. A laboratory test is the only way to be certain that the patient had that virus.
• Nigeria's numbers look to be rising alarmingly fast. Relative to each other they are, but compared to the dozens of new EVD cases being added between reports in other countries, it is still a small (although still very bad for Nigeria!) increase. This highlights that care is needed when reading charts. Perhaps also an understanding that between different outbreaks, the rate of new cases being added is disease specific. Lots of influenzavirus detections during flu season is what we expect, any ebolavirus cases are not what we expect nor what we want to see. Context. A hard thing to account for and probably a matter of experience.

Graph 2. The cumulative case graph. Adding new numbers to the sum of all the numbers that came before. 
Click on image to enlarge.

Graph 3. Changing the scale. Raising the primary y-axis (left) scale to 750, the level of the other country graphs, makes Nigeria's case numbers look tiny. But it underestimates the impact of the localised spread of Zaire ebolavirus in an are that was not part of the outbreak until a case flew in and spread it. Changing the scale is not just whimsical decision making, it can highlight the importance of events that may otherwise go unnoticed.
Click on image to enlarge.

Take care when interpreting a graph - look at the axes and also use your noodle

Finally, I'm going to look at the way in which I present the numbers I plot on a graph. I'm using the cumulative case chart for Liberia as my example (Graph 4 collection). Its the same one used in Graph 3 - the only thing different is that I've dragged the x-axis to the left (shrunk) or to the right (stretched) to see what that does. 

• The line plots look more or less steep when you shrink or stretch the x-axis, respectively. But the numbers have not changed. Possibly, our interpretation of them has, as a result of seeing the slope change. Remember though, check the axes. If you look at the x-axis, the shrunken version shows that those cases have climbed over a longer period than the slope suggests. Always check the denominator (the y of x/y) when you think about slope. Equally, the flatter curves of the stretched out x-axis, at the bottom of the Graph 4 collection, have to be looked at in context with time. The dates have been dragged out to what may be an unreasonable length, which makes the slopes look less; but they are still steeper in July than they were in April. Look around the graph for comparison. 
• As I said above, the current multi-country outbreak lets us compare and so we can see that some areas are adding new cases very rapidly between each report (Liberia and Sierra Leone) while others (Guinea) are not adding as many as quickly. Nigeria looks to have jumped quickly but that is also because of the altered scale (discussed above) 
• On VDU I get around this by also adding charts that plot total numbers per day or week or month or year. This shows a more discrete series of data that grow or shrink as the outbreak peaks or resolves. The 2 peaks of influenza A(H7N9) virus outbreaks illustrate this nicely - especially when combined with a cumulative case chart (Graph 5)!
• There is no real right or wrong here (although there are pixel width constraints)- but don't let your perceptions fool you when looking at someone's graphs for the first time. Take some time to really look at the graphs.

Graph 4 collection. Stretching the x-axis can seem like stretching the truth. But carefully read the axes. Some experience is needed here and ultimately you are at the mercy of the person presenting the data.
Click on image to enlarge.

Graph 5. Influenza A(H7N9) virus outbreak in China during 2013 and 2014. Plotting the numbers discretely (by week) clearly shows the two outbreak peaks (darker blue lines joining the data point dots) and gives valuable context to the cumulative graph in the background (pale blue mountain). This is probably my favourite style of disease numbers graph.
Click on image to enlarge.

I hope that has helped make sense of my graphs, and perhaps those of others too. I'm always on Twitter so hit me up with questions about this or requests for more posts like this, or to tell me whether it was helpful.

References

1. http://www.who.int/csr/disease/ebola/EVD_WestAfrica_WHO_RiskAssessment_20140624.pdf?ua=1

Now for something (not so) completely different: H7N9 maps...

Now it's time to mess around with influenza A(H7N9) virus mapping using Tableau.

I've (only just) realised the my esteemed peer, Shane Granger has been using Tableau to do this for ages (see here), and that this will be duplicating his excellent work. So I'll try my best to consciously differentiate my maps from his - but there's only so far you can go with that and there will be overlap. 

The page below is a very early first play with H7N9. It's just detections broken across 2013 and 2014, by province most likely to have been the source of the infections (as far as I can tell) in mainland China. 

If I can master this I'll try and add more details in the future. For now, these numbers a a little out of date but he trends are similar. This charts 449/452 detections.

Learn About Tableau

Snapdate: Avian influenza A(H7N9) virus...

There seem to have been more announcements of late than previously so I thought I'd plot this and see. 

These are a little adrift as the last 7 or so have not been through the WHO scrubbing process (which adds extra bits of data) so we will see a little shifting the last 2 or so blue dots on the chart below.

Click on image to enlarge.

Guangdong and Anhui provinces have the most active case generators in May.

Anhui province has reported 3 cases in a week and there seem to have been a constant stream of cases in May, but they they don't, in reality, seem to be out of what's become the ordinary in 2014 for a virus that is happily ticking over in several provinces.

Avian influenza A(H7N9) virus found in more than half of wet markets in Guangdong...

It comes as no surprise to me, but is still a very welcome piece of data, that Guangzhou's ongoing live bird markets and concurrent continued cases of H7N9 in people, are also happening in a an environment of 60% of market stalls tested positive for the virus in April.

A report in the South China Morning Post noted 

"Upon conclusion of the trial on September 30, the city government proposes gradually extending the ban, covering chickens, ducks, geese and pigeons, to other parts of the metropolis. The ban is expected to be implemented citywide by 2024."

"Currently, it affects 298 live poultry stalls at 82 wet markets in Yuexiu district, and in parts of Tianhe, Liwan and Panyu districts, where vendors will sell centrally slaughtered chickens that will be provided by three designated suppliers."

This is welcome news and a positive step towards stopping not just H7N9, but a raft of other influenza viruses that jump to us from, and mix to create new virus within, birds.

Source...

1. http://www.scmp.com/news/china/article/1505389/guangzhou-begins-trial-ban-live-poultry

H5N1 versus H7N9...

Data from the World Health Organization (2).

Green bars include surviving and fatal H5N1 laboratory-

confirmed cases in humans. The green "mountain" (area 

under the curve) is the accumulating tally of total cases. 

The red area-under-the-curve is the accumulating tally of 

fatal cases. The current total H7N9 cases is shown as a 

horizontal dashed blue line.

Click on image to enlarge.

This remains a kind of a pointless exercise. As I noted when I posted this first time back in February, but since I'm preparing some lectures I thought I'd post the latest version anyway.

These avian influenza A(H5N1) virus numbers have been curated since 2003 when the World Health Organization started an official tally. To that chart I've added where the current total number of laboratory confirmed human cases of infection by avian influenza A(H7N9) virus sits on the accumulating case tally (the green area-under-the-curve line). This blue dashed line highlights what we've heard before; H7N9 cases are piling up faster than H5N1 cases did. 

From 2003 it took H5N1 human cases nearly 6-years to reach the 430'ish mark; it's taken H7N9 about 61 weeks.

Sources...

1. Monthly risk assessment summary |  Influenza at the Human-Animal Interface
   http://www.who.int/influenza/human_animal_interface/HAI_Risk_Assessment/en/

H7N9 Snapdate: some quick charts...

Click on image to enlarge.

I don't have a lot of time tonight so this is just a quick post of some updated charts with a few summaries of some key features of the influenza A(H7N9) virus situation in south-eastern China. At writing it was at 432 detections with media reporting 128 deaths

Click on image to enlarge.

Guangdong is where H7N9 is still most active and it is this province that is the source of the continued cases trickling off Wave 2's peak.

Most H7N9 cases overall have been in Zhejiang and Guangdong provinces but lately, post-peak of Wave 2, there has been continued activity in Jiangsu province including a recent healthcare worker with no mention of "contact with poultry"; the absence of which stands out in World Health Organisation (WHO) reports because most cases are followed by affirmation of that phrase.

Click on image to enlarge.

In  the  survival chart above we see that most of the fatal cases, shown in red, are defined by an older age. Unfortunately, a lot more of the fatalities have been reported through the media without identifying details (48 of 128), than have come through official Chinese channels and out via the WHO. This lack of detail makes it impossible to clearly link a lot of the deaths to the case announcements. Only the custodians of these data know what this chart should really look like. NB: Since making the chart this morning I've found a handful more case details at FluTrackers, but public detail on fatal cases remains the weakest of any of the H7N9 data.

Click on image to enlarge.

We can see in the weekly chart on the right that the two H7N9 waves differed in timing, the width of their bases (more cases in Wave 2) as well as how "tight" their peaks were. Wave 2 has tailed off, but continues to spit out cases, while Wave 1 comprised both a steep climb and a steep decline in human cases.

Click on image to enlarge.

If we zoom in on Wave 2 we can see by looking at cases per day in the chart on the left, that between 0-4 illness onsets per day are being reported, as they have been since late Feb-2014. 

Is this the legacy of those regions whose live bird markets remained open or were only shut temporarily for disinfecting and restocking? Those regions with markets that were shut for much longer, or for good, do not seem to have contributed much to the continuing leak of H7N9 infections despite being key contributors during the peak periods before markets were closed.

Click on image to enlarge.

In zooming in on Wave 2's cases by week, but this time based on the region of likely acquisition of infection, we see that Guangdong province (brown line) has been the most consistent contributor of human H7N9 infections both late during the 2nd of the Wave 2 peaks, but also after the peak's decline almost everywhere else in south-east China. There was considerable publicised unwillingness from poultry producers to permanently close markets in this Province, a location with a major role in the nations poultry production. And so this little experiment incubates further and I have little doubt we will see the impact of that unwilingness late in 2014. 

Click on image to enlarge.

As noted above, public H7N9 death data do not allow good linkage with official case announcement data for about 48 fatalities, so my second-last chart tonight uses both public and media-release numbers to try and illustrate how the proportion of fatal cases (PFC) has changed across both Waves. The PFC seems to be holding fairly steady now between 17% and 30% (depending on source of numbers).

Click on image to enlarge.

And finally we see that the age and sex distribution across all cases (both Waves) is skewed to wards older males. Same as usual. If we look at this distribution (ran out of time to put in here) for the fatal cases, it is much more tightly grouped around the >60-year olds, but that females appear to dominate males in deaths during Wave 2, whereas it was the other way around for Wave 1.

Watching zoonoses evolve...

Special guest writer: @influenza_bio

For the first time in human history, we are watching diseases jump from animals to humans on a large scale. We've seen diseases appear for the first time in humans before; that's not new. We've seen HIV and several new strains of influenza emerge over the past century or so, for example. What is new is that we can now watch this process as it happens. We are able to watch animal diseases trickle case by case into humans, and we wonder whether any of these diseases might some day become human diseases. We wonder whether we might be watching pandemics develop in real time.

A disease that jumps from a non-human animal to a human (or the other way around) is called a "zoonotic" disease or a "zoonosis." Individual cases are called "zoonotic" cases. When a zoonotic disease is trying to make the jump to us permanent, we call this disease an "emerging infectious disease."

We have certainly been watching a lot of zoonotic MERS coronavirus and bird flu (e.g., H7N9 and H5N1) cases develop in people lately, along with Ebola virus cases. Zoonotic cases of other diseases, including infections with various strains of bird and swine flu, occasionally develop, as well, and are watched closely.

When the 2009 H1N1 flu pandemic started, we had no clue much beforehand that it was on its way.  We didn't even have surveillance data about swine flu strains that were even particularly close to the strain that emerged in us. A large animal flu surveillance gap blindsided us that year.

And we will undoubtedly be blindsided again by other emerging infectious diseases that we won't even see coming, although people are doing their best to see what's out there.

When an emerging infectious disease jumps to humans, it can cause either a relatively local outbreak or a worldwide outbreak, called a "pandemic." If a disease becomes a pandemic, that just means that it's spreading worldwide; the word "pandemic" doesn't imply anything about how bad the disease might or might not be. In some sense, the worst case can be when a disease jumps to humans and becomes "endemic" in humans, meaning that it gets established in people and regularly infects people, year after year. Endemic diseases can circulate worldwide (e.g., influenza) or in more restricted geographical regions (e.g., malaria).

Our knowledge and resources have grown to the extent that we are currently able to monitor some significant zoonotic outbreaks of disease. We are currently watching the MERS coronavirus and the influenza A(H7N9) virus both try to become human viruses.

Will either one succeed? We can't say. We've never watched this process happen before. We don't know how long such a process "usually" takes, or whether there even is a "usual" amount of time that it takes. We don't know how long it might take, or how quickly it has happened before. We doknow that the process is "stochastic," meaning that it involves a lot of chance. A pathogen that in one situation might cause a pandemic might just die out in another situation. Everything depends on the specific changes in a pathogen that get a chance to develop and on whether those changes end up getting passed on. We don't know how often pathogens "fail" when they "try" to make the jump to humans.

A lot of us have watched the recent surge in MERS coronavirus cases with some amount of concern. As of April 19, 2014, there are two large clusters of cases in the Middle East, and at least one of them is still growing. One cluster, in Jeddah, Saudi Arabia, now has 60 cases; 7 cases were added to this cluster today, and 6 were added yesterday. There are perhaps over a dozen cases in another cluster in the UAE. One patient who became ill with MERS in Jeddah at the end of March flew to his home country of Malaysia while ill and subsequently died in Malaysia; 79 of his contacts are now being watched closely in Malaysia. Test results are starting to come in for a number of these contacts, and thankfully all are negative for MERS so far. An asymptomatically infected Filipino health care worker traveled on an airplane back to the Philippines a few days ago. Yesterday, a MERS case was announced in Greece; a Greek man who had been living in Saudi Arabia was recently in Jeddah and presumably became infected there before flying back to Greece. He arrived in Greece with a fever; his contacts are now being monitored. In other words, MERS case numbers are growing quickly right now, at least in part through human-to-human transmission, and infected – and potentially infectious – patients are getting on airplanes to travel around the globe.

Does what we're seeing now represent changes in the virus that are making it more transmissible among humans? Or are we seeing a random fluctuation in the numbers of cases? Or, are we seeing more cases simply as a result of improved surveillance? I would argue that what we're seeing likely reflects one or more changes in the virus, simply because

1. We've been seeing so many more symptomatic cases recently, 
2. We've been seeing significantly larger clusters than we've ever observed before,
3. A greater number of health care workers appear to be getting infected than ever before, and
4. A greater proportion of cases are in health care workers than ever before. 

It's not that we've been seeing a rise only in the number of asymptomatic cases detected, which could suggest that we're only seeing the effects of improved surveillance. Moreover, while surveillance does seem to be picking up more mild and asymptomatic cases, it is difficult to know whether we are seeing more of these cases because of improved surveillance or because there simply are more such cases now. A lot of variables are being changed at the same time, and we don't have perfect information.

Nonetheless, the sheer numbers of recent cases suggest to me, at least, that the virus is changing and becoming more transmissible among humans. Until recently, we rarely saw evidence for human-to-human transmission of MERS; most cases may have been zoonotic. Now, however, large clusters involving roughly 1 to 4 dozen people are being seen, with single infected individuals infecting possibly up to a dozen or more other people. This is new. I don't think that we're seeing these clusters just as a result of improved surveillance, although I would be very happy to be wrong.

What does the future hold for MERS? We can't know. We might be watching MERS become a pandemic, and we might not. We might be watching the current relatively small MERS outbreak develop into a larger outbreak that eventually gets contained, as was seen with SARS. Or, the whole outbreak might all just simmer down or go away. Even if the virus were currently changing to become more transmissible, the current spate of cases could still simmer down or go away, just stochastically, just through sheer chance.

Prudence would dictate that we remain concerned and vigilant, however, especially as symptomatic MERS cases have had an approximately 40% case fatality rate (CFR). If MERS did cause one or more wider outbreaks in humans, that CFR might or might not change. Even if the CFR dropped to 10% of what it is now, it would still be on the same scale as the CFR for the 1918-19 influenza pandemic.

As a global society, we have an obligation to do everything in our power to prevent the MERS coronavirus from causing larger disease outbreaks in humans. We need more surveillance in affected countries, including much more genetic sequence data. And in countries of the Arabian Peninsula that are currently detecting MERS cases, infection control procedures need to be improved to the point where nosocomial cases in health care workers and patients are prevented. Health care workers in other countries should be educated about the possibility of MERS patients arriving from afar and about how to treat such patients safely. If this virus becomes more transmissible, we should not be caught unprepared. We can see this one coming.

Can we believe every H7N9 seroprevalence study we see?

Special Guest writer: @influenza_bio

A little over a year ago, the first known human patient got sick with avian influenza virus(H7N9). The number of H7N9 cases rose and fell in the spring of 2013, and a total of 134 people were known to have contracted H7N9 before June, 2013. Since then, sporadic cases appeared in the summer and fall, and by the end of December, 2013, new cases started to pick up again. We have now seen a second wave rise and fall, although several new cases still being reported each week. As of the time of this writing, just over 400 people are known to have been infected with H7N9. The case fatality rate (CFR) – roughly speaking, the percentage of people infected with H7N9 who die from it – for these known cases is almost 40%.

One question that is on a lot of people's minds is, how many other H7N9 cases are out there that we don't know about? How many mild cases are there that never get tested? How many asymptomatic cases are there that are missed? If there were a lot of undetected cases out there, that would mean that H7N9 is a lot less fatal than the known cases would make us think. On the other hand, if we were somehow miraculously seeing every single actual case, then the CFR would be as bad as all of these cases make it out to be. (And imagine what the CFR would be like without hospitals, ventilators and oseltamivir!)

How do we find out if there are cases that we're missing? One way is to do what is called a seroprevalence study. This means collecting blood samples from as wide a swath of a population as possible and testing to see how many of these samples have antibodies to H7N9. Antibodies are molecules that are made by cells of the immune system and that stick to specific pathogens to help our bodies to rid themselves of these pathogens. If someone gets sick with H7N9 influenza, his or her body would most likely continue to produce a significant amount of antibodies specifically against that strain for at least a good number of months after infection and possibly much longer. In general, people who are infected with influenza but who do not develop symptoms will also produce such antibodies, but their bodies will make fewer of them, and, on average, they won't make as many of them for as long. We don't know exactly what the pattern of antibody production is for people who are infected with H7N9 but don't develop symptoms, though, because researchers haven't identified enough of these individuals to study.

It is very important that we get these seroprevalence studies right. If they're done wrong and we miss a lot of cases, then we will simultaneously underestimate how common H7N9 cases are and overestimate how deadly the strain is. On the other hand, if seroprevalence studies are done wrong and we think a lot of people were infected with H7N9 when they weren't, then we will overestimate how common H7N9 cases are but underestimate how deadly the strain is. Facts can help us to respond to H7N9, and if we get the facts wrong, then we can't respond properly. For example, if we come to think mild H7N9 cases are far more numerous than the severe ones that actually get diagnosed, then we might not worry as much about H7N9 as we should.

What I'd like to talk about here are some of the important ways that seroprevalence studies can go wrong. To answer my title question, no, we cannot always believe the conclusions of every seroprevalence study we see. Scientists make mistakes, just like everyone else, and sometimes things just go wrong, too. I'd like for you to understand just how some of these mistakes can arise, so that you can better judge for yourself whether a study is likely to be reliable or not, or so that you can at least know that there are things out there that can go wrong.

How are seroprevalence studies done?

There are 2 types of laboratory assays (tests) that are usually used in seroprevalence studies (although there are others): hemagglutination inhibition (HI) assays¹and microneutralization (MN) assays.² (For more information about the HI assay in general, see a nice description by Dr. Racaniello.³) MN assays are considered better (more sensitive and specific) than HI assays, but they are harder to do. MN assays require a significant amount of extra work at the end that HI assays don't. But, more importantly for H7N9 studies, HI assays can be done with either "killed," modified or "live" virus, whereas MN assays require "live" H7N9 virus. In other words, HI assays can be done in almost any lab, but MN assays require a BSL-3 lab. A neutralization assay⁴ has been developed that uses a "pseudovirus" instead of live H7N9 and is therefore far less hazardous to work with, but formal WHO diagnostic criteria still require standard HI and/or MN assays.

First, blood samples are collected. Each blood sample is drawn into a tube, and after 15-30 minutes, the tube is centrifuged to separate clotted red blood cells from the rest of the blood. The red blood cells are discarded; what's left is called serum, and that's what's studied. The serum samples should then be put in a refrigerator if they'll be studied within a few days; if they'll be studied later, they should be frozen. Once a researcher is ready to study the serum samples, the serum samples are thawed. Virus is also used for the assay, so one or more tubes of virus are thawed, too. Different types of mammalian or bird cells are prepared: typically horse, turkey or chicken red blood cells for HI assays, or a special type of dog kidney cells ("MDCK" cells) for MN assays. Various solutions are prepared. Serum samples, virus preparations and cells are diluted as needed, and everything is transferred into little wells in a plastic "plate" in just the right way. In the HI assay, the plate then sits at room temperature for 1 hour, after which it is "read" by eye. In the MN assay, the plate then sits at 37°C (body temperature) for 19-21 hours, after which it is read by a machine (an "ELISA reader"). The assay is done. The results of the assay are then written down and analyzed, and voilà, a paper appears in the scientific literature.

What could possibly go wrong with these blood tests?

Let's start with some things that can go wrong with the lab work:

1. If blood samples are left sitting around for a long time without being centrifuged, the red blood cells will start to break apart, and enzymes released from the red blood cells will start to destroy antibodies (and everything else) in the blood samples. This happens even faster if blood samples are not refrigerated.
2. If serum samples are left in the fridge too long, things can start to deteriorate, just like food in your fridge would. The antibodies that you would like to measure start to be broken down. (Sometimes, for many different kinds of studies, people study serum samples left over after patients' blood tests at hospitals. Those samples sometimes sit around in a fridge for quite a while. Some of them can even be green from stuff growing in them while they're sitting around. Yuck.)
3. If plasma (what's left in blood after unclotted red blood cells are removed) is used instead of serum (what's left in blood after clotted red blood cells are removed), then the assay can read artificially high. Serum should always be used, not plasma.
4. Every time serum is frozen and thawed, some of the antibodies are effectively destroyed. This should not be done over and over. Serum samples should be put into the right size tubes that the researcher will want to use, so that the samples are put through only 1-2 "freeze-thaw cycles" before they are tested. And all serum samples should go through the same number of freeze-thaw cycles.
5. The same thing is true for virus samples used in MN assays. A single freeze-thaw cycle can reduce virus infectivity by a factor of 10. Virus samples also need to be kept on ice when they're being worked with.
6. The plate can be read wrong. It's hard to imagine reading an HI assay plate wrong, but a special procedure (ELISA) and special equipment (ELISA plate reader) are used in the MN assay, and ELISA assays can go wrong.

But, hopefully all of that was done right. Not all researchers, students and technicians are created equal, but hopefully the lab "PI" (Principal Investigator; the person running the lab) is competent and ensures that everyone is doing things correctly.

What could go wrong with the data analysis?

What else could go wrong? The data analysis might not be done correctly. And it's here where perfectly good data can be ruined and where you have to look at seroprevalence studies most closely.

Suppose you've measured your antibody amounts ("titers") in your serum samples. How do you decide which titers mean the sample came from someone who was infected with H7N9, and which titers mean they didn't? Do you just pick a number out of thin air? If you don't have data to tell you which titers mean what, then all you are doing is measuring antibody levels in a population, and you can make no interpretation about what those levels mean. You can't say that they mean any people have or have not been infected with H7N9 at all.

Instead, you need actual measurements using serum samples from people who are known to have been infected with H7N9 to tell you what your titers mean. Someone has to study a number of patients to see what their actual H7N9 antibody titers are, and then a mathematical analysis of that data is done to come up with a threshold titer value, above which serum samples can be said to have come from people infected with H7N9 with some large degree of certainty, and below which they are thought to have come from people who were not infected. We've seen almost no asymptomatic cases (cases with no symptoms), so we really can't say much about them. So we have to go with data from H7N9 patients who have had symptoms. Here's a great graph showing antibody titers, as measured using the HI assay, in serum samples from H7N9 patients:⁵

Figure 1. H7N9 HI
Euro Surveill. 2013 Dec 12;18(50):20657

As you can see in the graph above (Figure 1), by around 3 weeks after infection onset, all samples from patients whose HI titer was measured had titers ≥40.

The graph below (Figure 2), from a different study,⁴ shows that the HI titer for all H7N9 samples studied by this set of authors was also ≥40. In addition, this graph shows titers from "control" samples (i.e., samples from people who did not have H7N9 infections); all control samples had titers that were <40.

Figure 2: H7N9 IC50 HI4
Emerg Infect Dis. 2013 Oct;19(10):1685-7

Finally, below (Figure 3) is another nice graph, from a third study,⁶ showing anti-H7N9 antibody levels ("IgG"), "HI" assay results and MN assay ("NAb") results for several H7N9 patients, again showing that all samples from the H7N9 patients studied had HI titers ≥40. This graph also shows that all H7N9 patient serum samples had an MN titer of ≥20, if samples were taken after enough time had elapsed since their infections had started.

Figure 3. H7N9 IgG HI NAb.
Emerg Infect Dis. 2014 Feb;20(2):192-200

In other words, if an individual's anti-H7N9 antibody titer is ≥40 by the HI assay or ≥20 by the MN assay, these data suggest that we could pretty safely say that he or she has had a symptomatic H7N9 infection within the past few months, and if the HI or MN titers are below those cutoffs, then the individual probably hasn't had a symptomatic H7N9 infection. We don't know to what extent asymptomatic H7N9 infections will be captured by these cutoffs, but it is likely that some asymptomatic cases would be missed using these cutoffs. It is also possible that some mild infections could be missed using these cutoffs. However, it would be a great step forward just to get estimates of what percentages of any regional population or occupational group of people have had any kind of H7N9 infection. A comparison of antibody titers for asymptomatically infected and symptomatically infected H5N1 cases may be instructive when thinking about H7N9.⁷

WHO guidelines are even stricter than the cutoffs discussed in the paragraph above. WHO guidelines say that, using the HI assay, only single samples with titers of ≥160 can be considered "seropositive": "Paired sera (acute and convalescent sera) with a 4-fold rise in HI titer or single sera collected in convalescent phase with HI titer of ≥160 could be considered as H7N9 HI antibody positive. Sera with HI titer of 20-80 should be confirmed by MN or WB assay."¹ For the MN assay, however, the WHO does not give specific cutoffs: "With single-serum samples, care must be taken in interpreting low titers such as 20 and 40. Generally, knowledge of the antibody titers in an age-matched control population is needed to determine the minimum titer that is indicative of a specific antibody response to the virus used in the assay."²

Now, it should be noted that WHO assay instructions recommend the use of horse red blood cells for the HI assay, and not everyone uses horse red blood cells. Some people use chicken, turkey, guinea pig or other kinds of red blood cells. That starts making comparisons between different groups' assays difficult. Horse red blood cells are better to use than turkey red blood cells for H7N9 because they have more a2,3-linked ("bird") sialic acids (influenza receptors); HI results are more sensitive with horse red blood cells. In other words, it may take less antibody in the assays to get the same result using horse red blood cells than it would using turkey red blood cells. This would translate into a higher number, when discussing H7N9 patient titers, for HI assays using horse red blood cells, compared to assays using turkey red blood cells. I have not seen direct comparisons of titers obtained using different types of red blood cells in HI assays specifically for H7N9, but the situation is probably similar to that for H5N1.⁸

Figures 1 and 3 above were made with HI data obtained using horse red blood cells. Figure 2 used guinea pig red blood cells. Are they completely comparable? No. Are they pretty comparable? Yes.

Are you getting a feeling for how complicated it is to interpret a seroprevalence paper? And for how difficult it is to compare results across studies?

Why does all of this matter?

It matters because some seroprevalence studies don't use appropriate cutoffs. And because it can be hard to determine even what an appropriate cutoff is when red blood cells from different species are used in an HI assay. This is where the reader has to be really careful. Cutoffs for seropositivity have been a big issue⁹ with H5N1 seroprevalence studies; some researchers have used cutoffs that were too low, and hence they have almost certainly overestimated how common H5N1-specific antibodies were in the populations studied.

So far, only one H7N9 serology paper published to date has reported probable seropositive samples, and this paper simply reported HI titers without using any specific threshold for seropositivity. Only one used study horse red blood cells in HI assays. The one paper that used an MN assay did use appropriate cutoffs. It should be noted that the new WHO HI guidelines were only published in December, 2013, after a couple of these papers were already published.

Here are the studies that have been published so far (I hope I haven't left any out):

1. Bai et al.¹⁰ looked at serum samples collected before November, 2012 from poultry workers in eastern China and found no H7N9-positive samples. The study used HI and MN assays. Turkey red blood cells were used in the HI assay. Appropriate cutoffs were used for the MN assay.
2. Hsieh et al.¹¹ studied 14 close contacts of the first H7N9 case in Taiwan. The authors took blood samples within 18-28 days after the contacts' earliest exposures. The authors used an HI assay but not an MN assay. They used turkey red blood cells for the HI assay. They found all contacts to have an HI titer £10, and declared all to be seronegative. The HI titer for the H7N9 patient in their study was 1:80. These conclusions seem very sound.
3. Yang et al.¹² looked at serum samples from 1129 people from regions of China in which H7N9 cases had been seen, and from 396 poultry workers from 10 districts in which H7N9 cases had been seen. None of the samples from the general population was found to be seropositive, whereas >6% of the poultry workers were found to be seropositive. The authors also examined serum samples from several H7N9 patients. The study used an HI assay but not an MN assay. The authors used a cutoff of ≥80, along with turkey red blood cells, for the HI assay. Because the authors examined serum samples from H7N9 patients using their methods and got results that are reasonably similar to other results, their cutoffs are most likely reasonable, and their conclusions are probably quite sound. The authors report:
• "Of the 1129 serum samples collected from individuals (age range, 1–88 years) in the general population, 9 (0.8%) had an HI titer of≥40 to influenza A(H7N9), but no serum samples with an HI titer of≥80 were found (Table 1). In contrast, among poultry workers, 13.9% (55/396) and 6.3% (25/396) had influenza A(H7N9) antibody titers of ≥40 and ≥80 (20 had an HI titer of 80, and 5 had an HI titer of 160), respectively."
  
  It is hard to imagine that an HI titer of 160 can be a spurious finding ("non-specific," to the initiated). Thus, these data strongly suggest that at least some H7N9 cases have been going undetected among poultry workers. Suppose we consider only the poultry workers with HI titer ≥80, or 6.3% of the poultry workers. If we then consider how many poultry workers there are, total, in districts from which H7N9 cases have emerged, then this study suggests that it's possible that quite a large number of poultry workers have been exposed to H7N9. Still, this study examined only a very small number of people, and we should be cautious about reading too much into these results.
4. Qiu et al.¹³ looked at 3 H7N9 patients and 3 close household contacts of the patients who were exposed before infection control practices were put in place. The authors looked for viral RNA using a sensitive test (PCR) and examined serum samples drawn 15-26 days post-exposure using both an HI assay and a pseudovirus-based neutralization assay. They found no contacts to be seropositive. The H7N9 patients had HI titers that reached 160-640 during this time, and the patient contacts all had titers <10. The authors used horse red blood cells for the HI assay. These findings also seem sound.

To summarize, the conclusions from all of these papers do seem sound. But, it would be wise to keep all of these issues in mind as subsequent studies appear over time.

An additional study¹⁴looked at antibody titers in 1723 serum samples collected in Vietnam using a very different kind of assay (a protein microarray). Because seropositivity cutoff levels had not been determined with authors' assay methods using actual H7N9 patient samples, these authors were appropriately very careful not to attempt to draw any conclusions about H7N9 seroprevalence from their data:

"Because titers calculated from our assay are not directly comparable to HI or microneutralization tests, no cutoff is chosen to represent positivity or clinical protection. It is not possible to associate these titers with past exposure or past infection, as serological assays have not yet been validated for H7N9."

For the future

So, as new H7N9 serology studies gradually come out, you be the judge. Figure out whether they're believable or not. Ask yourself the following:

1. What assay(s) were used? Did the authors use an MN assay? They get bonus points if they did. 
• If only an HI assay was used, then the conclusions are slightly less certain than if an MN assay was used.
2. If the authors used an HI assay, what species were the red blood cells from?
• If horse red blood cells weren't used, then HI titer cutoffs lower than 160 are probably appropriate, but there is also more uncertainty about what an appropriate cutoff would be.
3. What cutoff(s) did they use for seropositivity in their assay(s)? Do these cutoffs mesh with WHO guidelines? Do they mesh with what we know about H7N9 patient HI and MN antibody titers?

References

1. http://www.who.int/influenza/gisrs_laboratory/cnic_serological_diagnosis_hai_a_h7n9_20131220.pdf
2. http://www.who.int/influenza/gisrs_laboratory/cnic_serological_diagnosis_microneutralization_a_h7n9.pdf
3. http://www.virology.ws/2009/05/27/influenza-hemagglutination-inhibition-assay/
4. Qiu C, Huang Y, Zhang A, Tian D, Wan Y, Zhang X, Zhang W, Zhang Z, Yuan Z, Hu Y, Zhang X, Xu J. Safe pseudovirus-based assay for neutralization antibodies against influenza A(H7N9) virus. Emerg Infect Dis. 2013 Oct;19(10):1685-7
5. Zhang A, Huang Y, Tian D, Lau EH, Wan Y, Liu X, Dong Y, Song Z, Zhang X, Zhang J, Bao M, Zhou M, Yuan S, Sun J, Zhu Z, Hu Y, Chen L, Leung CY, Wu JT, Zhang Z, Zhang X, Peiris JS, Xu J. Kinetics of serological responses in influenza A(H7N9)-infected patients correlate with clinical outcome in China, 2013. Euro Surveill. 2013 Dec 12;18(50):20657 
6. Guo L, Zhang X, Ren L, Yu X, Chen L, Zhou H, Gao X, Teng Z, Li J, Hu J, Wu C, Xiao X, Zhu Y, Wang Q, Pang X, Jin Q, Wu F, Wang J. Human antibody responses to avian influenza A(H7N9) virus, 2013. Emerg Infect Dis. 2014 Feb;20(2):192-200
7. Buchy P et al., PLoS One. 2010 May 27;5(5):e10864
8. See, e.g., Table 4 in Pawar SD et al., Virol J. 2012 Oct 30;9:251
9. Osterholm MT and Kelley NS, MBio. 2012 Feb 24;3(2):e00045-12
10. Bai T et al., N Engl J Med. 2013 Jun 13;368(24):2339-40
11. Hsieh SM et al., J Infect. 2013 Nov;67(5):494-5
12. Yang S et al., J Infect Dis. 2014 Jan 15;209(2):265-9
13. Qiu C et al., J Clin Virol. 2014 Feb;59(2):129-31
14. Boni MF et al., J Infect Dis. 2013 Aug 15;208(4):554-8

NOTE: I did not have a hand in writing this post and thus take no credit for it. This was entirely the work of the Guest Writer.