Photo: M.B.M./Unsplash

Imagine a science textbook without images. No charts, no graphs, no illustrations or diagrams with arrows and labels. The science would be a lot harder to understand.

Thats because humans are visual creatures by nature. People absorb information in graphic form that would elude them in words. Images are effective for all kinds of storytelling, especially when the story is complicated, as it so often is with science. Scientific visuals can be essential for analyzing data, communicating experimental results and even for making surprising discoveries.

Visualisations can reveal patterns, trends and connections in data that are difficult or impossible to find any other way, says Bang Wong, creative director of MITs Broad Institute. Plotting the data allows us to see the underlying structure of the data that you wouldnt otherwise see if youre looking at a table.

And yet few scientists take the same amount of care with visuals as they do with generating data or writing about it. The graphs and diagrams that accompany most scientific publications tend to be the last things researchers do, says data visualisation scientist Sen ODonoghue. Visualisation is seen as really just kind of an icing on the cake.

As a result, science is littered with poor data visualisations that confound readers and can even mislead the scientists who make them. Deficient data visuals can reduce the quality and impede the progress of scientific research. And with more and more scientific images making their way into the news and onto social media illustrating everything from climate change to disease outbreaks the potential is high for bad visuals to impair public understanding of science.

The problem has become more acute with the ever-increasing amount and complexity of scientific data. Visualisation of those data to understand as well as to share them is more important than ever. Yet scientists receive very little visualisation training. The community hasnt by and large recognised that this is something that really is needed, says ODonoghue, of the University of New South Wales and lead author of a paper about biomedical data visualisation in the 2018 Annual Review of Biomedical Data Science.

There are signs of progress, however. At least two annual conferences dedicated to scientific data visualisation have sprung up in the last decade. And the journal Nature Methods ran a regular column from 2010 to 2016 about creating better figures and graphs, which was then adapted into guidelines for scientists submitting papers to that journal. But so far, few scientists are focusing on the problem.

Improving scientific visualisation will require better understanding of the strengths, weaknesses and biases of how the human brain perceives the world. Fortunately, research has begun to reveal how people read, and misread, different kinds of visualisations and which types of charts are most effective and easiest to decipher. Applying that knowledge should lead to better visual communication of science.

We have a lot of practical knowledge about what works and what doesnt, says computer scientist Miriah Meyer of the University of Utah. There are a lot of principles that have been through the test of time and have been shown over and over again to be really effective.

Chart choice

The human visual system evolved to help us survive and thrive in the natural world, not to read graphs. Our brains interpret what our eyes see in ways that can help us find edible plants among the toxic varieties, spot prey animals and see reasonably well in both broad daylight and at night. By analyzing the information we receive through our eyes to serve these purposes, our brains give us a tailored perception of the world.

In the early 1980s, Bell Labs statisticians William Cleveland and Robert McGill began researching how the particulars of human perception affect our ability to decipher graphic displays of data to discover which kinds of charts play to our strengths and which ones we struggle with. In a seminal paper published in 1984 in the Journal of the American Statistical Association, Cleveland and McGill presented a ranking of visual elements according to how easily and accurately people read them.

Their experiments showed that people are best at reading charts based on the lengths of bars or lines, such as in a standard bar chart. These visualisations are the best choice when its important to accurately discern small differences between values.

Study participants found it somewhat harder to judge differences in direction, angle and area. Figures using volume, curvature or shading to represent data were even tougher. And the least effective method of all was colour saturation.

The ability of the audience to perceive minute differences is going to get worse and worse as you move down the list, says computer scientist Jeffrey Heer of the University of Washington in Seattle. In general, its best practice to use the highest graphical element on the list that meets the needs of each type of data.

For example, if its important to show that one particular disease is far more lethal than others, a graphic using the sise of circles to represent the numbers of deaths will do fine. But to emphasise much smaller differences in the numbers of deaths among the less-lethal diseases, a bar chart will be far more effective.

In 2010, Heer used Amazons Mechanical Turk crowdsourcing service to confirm that Cleveland and McGills ranking holds true in the modern digital environment. Since then, Heer, ODonoghue and others have used crowdsourcing to test many other aspects of visualisation to find out what works best. That has huge power going forward to take this whole field and really give it a solid engineering basis, ODonoghue says.

Pernicious pies

Cleveland and McGills graphical ranking highlights why some popular types of figures arent very effective. A good example is the ever-popular pie chart, which has earned the disdain of data visualisation experts like Edward Tufte. In his influential 1983 treatise, The Visual Display of Quantitative Information, Tufte wrote that the only design worse than a pie chart is several of them.

Pie charts are often used to compare parts of a whole, a cognitively challenging visual task. The reader needs to judge either differences between the areas of the pie slices, or between the angles at the center of the chart: Both types of estimations are more difficult than discerning the difference in lengths of bars on a bar chart, which would be a better option in many instances.

Pie charts can be tempting because they are generally more attractive than bar charts, are easy to fill with colours and are simple to make. But they are rarely the best choice and are acceptable only in limited contexts. If the goal is to show that the parts add up to a whole, or to compare the parts with that whole (rather than comparing slices with each other), a well-executed pie chart might suffice as long as precision isnt crucial.

For example, a pie chart that depicts how much each economic sector contributes to greenhouse gas emissions nicely shows that around half come from electricity and heat production along with agriculture, forestry and other land use. Transportation, which often gets the most attention, makes up a much smaller piece of the pie. Putting six bars next to each other in this case doesnt immediately show that the parts add up to 100 percent or what proportion of the whole each bar represents.

In some scientific disciplines, the pie chart is simply standard practice for displaying specific types of data. And its hard to buck tradition. There are certain areas in epigenetics where we have to show the pie chart, says Wong, who works with biomedical scientists at the Broad Institute to create better visualisations. I know the shortcomings of a pie chart, but its always been shown as a pie chart in every publication, so people hold on to that very tight.

In other instances, the extra work pies ask of the human brain makes them poor vehicles for delivering accurate information or a coherent story.

Behind bars

Though bar graphs are easy to read and understand, that doesnt mean theyre always the best choice. In some fields, such as psychology, medicine and physiology, bar graphs can often misrepresent the underlying data and mask important details.

Bar graphs are something that you should use if you are visualising counts or proportions, says Tracey Weissgerber, a physiologist at the Mayo Clinic in Rochester, Minnesota, who studies how research is done and reported. But theyre not a very effective strategy for visualising continuous data.

Weissgerber conducted a survey of top physiology journals in 2015 and found that some 85% of papers contained at least one bar graph representing continuous data things like measurements of blood pressure or temperature where each sample can have any value within the relevant range. But bars representing continuous data can fail to show some significant information, such as how many samples are represented by each bar and whether there are subgroups within a bar.

For example, Weissgerber notes that the pregnancy complication preeclampsia can stem from problems with the mother or from problems with the baby or placenta. But within those groups are subgroups of patients who arrive at the same symptoms through different pathways. Were really focused on trying to understand and identify women with different subtypes of preeclampsia, Weissgerber says. And one of the problems with that is if were presenting all of our data in a bar graph, there are no subgroups in a bar graph.

Bar charts are especially problematic for studies with small sample sizes, which are common in the basic biomedical sciences. Bars dont show how small the sample sizes are, and outliers can have a big effect on the mean indicated by the height of a bar.

One of the challenges is that in many areas of the basic biomedical sciences, bar graphs are just accepted as how we show continuous data, Weissgerber says.

There are several good alternative graphs for small continuous data sets. Scatterplots, box plots and histograms all reveal the distribution of the data, but they were rarely used in the papers Weissgerber analysed. To help correct this problem, she has developed tools to create simple scatterplots and various kinds of interactive graphs.

Ruinous rainbows

Colour can be very effective for highlighting different aspects of data and adding some life to scientific figures. But its also one of the easiest ways to go wrong. Human perception of colour isnt straightforward, and most scientists dont take the peculiarities of the visual system into account when choosing colours to represent their data.

One of the most common bad practices is using the rainbow colour scale. From geology to climatology to molecular biology, researchers gravitate toward mapping their data with the help of Roy G. Biv. But the rainbow palette has several serious drawbacks and very little to recommend it.

Even though its derived from the natural light spectrum, the order of colours in the rainbow is not intuitive, says Wong. You sort of have to think, is blue bigger than green? Is yellow larger than red?

An even bigger problem is that the rainbow is perceived unevenly by the human brain. People see colour in terms of hue (such as red or blue), saturation (intensity of the colour) and lightness (how much white or black is mixed in). Human brains rely most heavily on lightness to interpret shapes and depth and therefore tend to see the brightest colours as representing peaks and darker colors as valleys. But the brightest colour in the rainbow is yellow, which is usually found somewhere in the middle of the scale, leading viewers to see high points in the wrong places.

Compounding the problem, the transitions between some colours appear gradual, while other changes seem much more abrupt. The underlying data, on the other hand, usually have a consistent rate of change that doesnt match the perceived unevenness of the rainbow. You can have perceptual boundaries where none exist and also hide boundaries that do exist, says climate scientist Ed Hawkins of the University of Reading in England. Even scientists can fall prey to this illusion when interpreting their own data.

To avoid the rainbow problem, some researchers have come up with mathematically based palettes that better match the perceptual change in their colours to changes in the corresponding data. Some of these newer colour scales work specifically for people with red-green colour blindness, which is estimated to affect around 8 percent of men (but only a tiny fraction of women).

Though cartographers and a few scientists like Hawkins have been railing against the rainbow for decades, it remains pervasive in the scientific literature. Some fields of science have probably been using it ever since colour printing was invented. And because many scientists arent aware of the problematic aspects of the rainbow, they see no reason to defy tradition. People are used to using it, so they like it, they feel comfortable with it, Hawkins says.

This inclination is also encouraged by the fact that the rainbow colour scale is the default for much of the software scientists use to create visualisations. But Hawkins and others have been pushing software makers to change their defaults, with some success.

In 2014, MathWorks switched the default for the MATLAB software program to an improved colour scheme called parula. In 2015 a cognitive scientist and a data scientist developed a new default colour scheme called viridis for making plots with the popular Python programming language. And a new mathematically derived colour scheme called cividis has already been added to a dozen software libraries, though it is not yet the default on any of them.

Also read: The Growth of Acronyms in the Scientific Literature

Hazardous heat maps

One of the most interesting quirks of the human visual system and one of the most nettlesome for data visualisation is that our perception of a colour can be influenced by other nearby colours. In some cases the effect is quite dramatic, leading to all sorts of optical illusions.

Whenever a visualisation places different colours, or even shades of the same colour, next to each other, they can interact in unintended ways. The exact same colour will look entirely different when surrounded by a darker shade than it looks when surrounded by a lighter shade, a phenomenon known as simultaneous contrast. A well-known illustration of this, the checker shadow illusion, plays with the brains interpretation of colours when a shadow is cast across a checkered grid.

The effect of colour interactions poses a huge problem, Wong says. In the life sciences, one pervasive example is the heat map, which is often used to reveal relationships between two sets of data. If you flip through a journal, a third of the figures are heat maps, he says. This is a very popular form of data visualisation that in fact is biasing scientific data.

A heat map is a two-dimensional matrix, basically a table or grid, that uses colour for each square in the grid to represent the values of the underlying data. Lighter and darker shades of one or more hues indicate lower or higher values. Heat maps are especially popular for displaying data on gene activity, helping researchers identify patterns of genes that are more or less actively producing proteins (or other molecules) in different situations.

Heat maps are great for packing a ton of data into a compact display. But putting various shades of colours right next to each other can trigger the simultaneous contrast illusion. For example, a scientist comparing the colours of individual squares in the grid can easily misinterpret two different shades of orange as being the same or think that two identical shades are quite different depending on the colours of the surrounding squares.

This is a huge problem in heat maps where youre relying on a bunch of colour tiles sitting next to each other, Wong says. This unintentional bias is sort of rampant in every heat map.

For gene activity data, green and red are often used to show which genes are more or less active. A particular shade of green can look very different surrounded by lighter shades of green compared with when it is surrounded by darker shades of green, or by red or black. The value that the shade of green is representing is the same, but it will appear higher or lower depending on its neighboring squares.

A blob of bright green squares in one part of the grid might mean that a gene is highly active in a group of closely related subspecies, say of bacteria. At the same time in another part of the grid, a single dull-green square surrounded by black squares may look bright, making it appear that the same gene is highly active in an unrelated bacterium species, when in fact it is only weakly active.

One way to mitigate the problem, Wong says, is to introduce some white space between parts of the grid, perhaps to separate groups of related species, groups of samples or sets of related genes. Breaking up the squares will reduce the interference from neighboring colours. Another solution is to use an entirely different display, such as a graph that uses lines to connect highly active genes, or a series of graphs that represent change in gene activity over time or between two experimental states.

Muddled messaging

Making sure a visualisation wont mispresent data or mislead readers is essential in sharing scientific results. But its also important to consider whether a figure is truly drawing attention to the most relevant information and not distracting readers.

For example, the distribution of many data sets when plotted as a line graph or a histogram will have a bell shape with the bulk of the data near the center. But often we care about whats on the tails, Wong says. For the viewer, thats often overwhelmed by this big old thing in the middle.

The solution could be to use something other than height to represent the distribution of the data. One option is a bar code plot, which displays each value as a line. On this type of graph, it is easier to see details in areas of low concentration that tend to all but disappear on a bell curve.

Thoughtfully applied colour can also enhance and clarify a graphics message. On a scatterplot that uses different colours to identify categories of data, for instance, the most important information should be represented by the colours that stand out most. Graphing programs may just randomly assign red to the control group because its the first column of data, while the interesting mutant that is central to the findings ends up coloured gray.

Pure colours are uncommon in nature, so limit them to highlight whatever is important in your graphics, writes data visualisation journalist Alberto Cairo in his 2013 book The Functional Art. Use subdued hues grays, light blues and greens for everything else.

Besides the rainbow and simultaneous contrast, there are plenty of other ways to get into trouble with colour. Using too many colours can distract from a visualisations main message. Colours that are too similar to each other or to the background colour of an image can be hard to decipher.

Colours that go against cultural expectations can also affect how well a reader can understand a figure. On maps that show terrain, for example, the expectation is that vegetation is green, dry areas are brown, higher elevations are white, cities are gray, and of course water is blue. A map that doesnt observe these well-established colour schemes would be much harder to read. Imagine a US electoral map with Democratic areas shown in red and Republican areas in blue, or a bar chart showing different causes of death in bright, cheery colours the dissonance would make it harder to absorb their message.

If colour isnt necessary, sometimes its safest to stick with shades of gray. As Tufte put it in his 1990 book Envisioning Information, Avoiding catastrophe becomes the first principle in bringing color to information: Above all, do no harm.

Visualise the future

Many data visualisation problems persist because scientists simply arent aware of them or arent convinced that better figures are worth the extra effort, ODonoghue says.

Hes been working to change this situation by initiating and chairing the annual Vizbi conference focused on visualising biological science, teaching a visualisation workshop for scientists, and combing the literature for evidence of the best and worst practices, which are compiled into his 2018 Annual Reviews paper. But overall, he says, the effort hasnt gained a lot of momentum yet. I think weve got a long ways to go.

One reason for the lack of awareness is that most scientists dont get any training in data visualisation. Its rarely required of science graduate students, and most institutions dont offer classes designed on scientific visualisation. For many students, particularly in the biomedical sciences, their only exposure to data visualisation is in statistics courses that arent tailored to their needs, Weissgerber says.

Scientists also tend to follow convention when it comes to how they display data, which perpetuates bad practices.

One way to combat the power of precedent is by incorporating better design principles into the tools scientists use to plot their data (such as the software tools that have already switched from the rainbow default to more perceptually even palettes). Most scientists arent going to learn better visualisation practices, ODonoghue says, but theyre going to use tools. And if those tools have better principles in them, then just by default they will [apply those].

Scientific publishers could also help, he says. I think the journals can play a role by setting standards. Early-career scientists take their cues from more experienced colleagues and from published papers. Some journals, including PLoS Biology, eLife and Nature Biomedical Engineering have already responded to Weissgerbers 2015 work on bar graphs. In the time since the paper was published, a number of journals have changed their policies to ban or discourage the use of bar graphs for continuous data, particularly for small data sets, she says.

With scientific data becoming increasingly complex, scientists will need to continue developing new kinds of visualisations to handle that complexity. To make those visualisations effective for both scientists and the general public data visualisation designers will have to apply the best research on humans visual processing in order to work with the brain, rather than against it.

This article originally appeared in Knowable Magazine, an independent journalistic endeavor from Annual Reviews.

Go here to see the original:

Why Scientists Need To Be Better at Visualising Data - The Wire Science

DANVILLE Among people who have strokes and COVID-19, there is a higher incidence of severe stroke as well as stroke in younger people, according to new data from a multinational study group on COVID-19 and stroke, led by a team of Geisinger researchers.

The COVID-19 Stroke Study Groups latest report, published in the journal Stroke, focused on a group of 432 patients from 17 countries diagnosed with COVID-19 and stroke. Among this group, the study found a significantly higher incidence of large vessel occlusion (LVO) strokes caused by a blockage in one of the brains major arteries that are typically associated with more severe symptoms. Nearly 45% of strokes in the study group were LVOs; in the general population, 24 to 38% of ischemic strokes are LVOs.

The study group also had a high percentage of young patients who had strokes: more than a third were younger than 55, and nearly half were younger than 65. Pre-pandemic general population data showed 13% of strokes occurred in people under 55, and 21% in people younger than 65.

The data showed that that less-severe strokes, mostly in critically ill patients or overwhelmed health centers, were underdiagnosed. This finding is significant, the research team said, as minor or less-severe stroke may be an important risk factor for a more severe stroke in the future.

Our observation of a higher median stroke severity in countries with lower healthcare spending may reflect a lower capacity for the diagnosis of mild stroke in patients during the pandemic, but this may also indicate that patients with mild stroke symptoms refused to present to the hospitals, said Ramin Zand, M.D., a vascular neurologist and clinician-scientist at Geisinger and leader of the study group.

Throughout the pandemic, people with COVID-19 have reported symptoms involving the nervous system, ranging from a loss of smell or taste to more severe and life-threatening conditions such as altered mental state, meningitis and stroke. A group of Geisinger scientists and a team of experts from around the world formed the COVID-19 Stroke Study Group shortly after the pandemic began to study the correlation between COVID-19 infection and stroke risk.

Results from the first phase of the study, which included data on 26,175 patients, indicated an overall stroke risk of 0.5% to 1.2% among hospitalized patients with COVID-19 infection. The finding demonstrated that, even though there were increasing reports of patients with COVID-19 experiencing stroke, the overall risk is low.

Our initial data showed that the overall incidence of stroke was low among patients with COVID-19, and while that hasnt changed, this new data shows that there are certain groups of patients for example, younger patients who are more affected, said Vida Abedi, Ph.D., a scientist in the department of molecular and functional genomics at Geisinger. We hope these findings highlight new research directions to better identify patients at risk and help improve the quality of care.

Geisinger has an exciting research environment with more than 50 full-time research faculty and more than 30 clinician scientists. Areas of expertise include precision health, genomics, informatics, data science, implementation science, outcomes research, health services research, bioethics and clinical trials.

Today's breaking news and more in your inbox

Read the original post:

Study: COVID- positive people have more severe strokes - Lock Haven Express

Land a career in Adobe with these top big data/data science jobs.

Many businesses encountered turbulence in 2020, yet big data/data science saw substantial demand and growth.

Data science professionals are in high demand all across the world. These job opportunities will continue to grow after 2021, with over 1.5 lakh more positions being added. This is a natural reaction to datas importance as a resource for businesses in the digital age. Weve compiled a list of the top 10 Big Data/Data Science job openings in Adobe to watch out for this month.

Location: Bangalore

Requirements:

Know more here.

Location: Noida, Uttar Pradesh

Responsibilities:

Know more here.

Location: Bengaluru, Karnataka

Responsibilities:

Know more here.

Location: Bengaluru, Karnataka

Responsibilities:

Know more here.

Location: Noida, Uttar Pradesh

Responsibilities:

Know more here.

Location: San Francisco

Responsibilities:

Know more here.

Location: San Jose

Responsibilities:

Know more here.

Location: Bangalore

Responsibilities:

Know more here.

Location: Bucharest

Responsibilities:

Location: San Jose

Responsibilities:

Know more here.

Share This ArticleDo the sharing thingy

About AuthorMore info about author

See the original post here:

Top Big Data/Data Science Job Openings in Adobe to Watch Out for This Month - Analytics Insight

Mozilla announced that Pyodide, a project aiming at providing a full Python data science stack running entirely in the browser, has become an independent community-driven project. Pyodide leverages the CPython 3.8 interpreter compiled to WebAssembly, and thus allows using Python, NumPy, Pandas, Matplotlib, SciPy, and more in Iodide, an experimental interactive scientific computing environment for the web.

The Pyodide team originally explained the rationale behind Pyodide as follows:

When we started thinking about making the web better for scientists, we focused on ways that we could make working with Javascript better, like compiling existing scientific libraries to WebAssembly and wrapping them in easy-to-use JS APIs. [] Mozillas WebAssembly wizards offered a more ambitious idea: if many scientists prefer Python, meet them where they are by compiling the Python science stack to run in WebAssembly.

The Iodide playground showcases a notebook that uses Python and Python packages in a JavaScript environment and vice versa:

Pyodide may be used in any context where it is necessary to run Python inside a web browser with full access to the Web APIs. The latest release note states that Pyodide converted the Python 3.8 runtime to WebAssembly, along with the Python scientific stack including NumPy (scientific computing), Pandas (data analysis), Matplotlib (plotting library), SciPy (scientific and technical computing), and scikit-learn (machine learning). 75 packages are available at the time of the release. Pure Python wheels may also be installed from the PyPi Python package manager. Python 0.17 additionally provides transparent conversion of objects between JavaScript and Python.

Iodide was created in 2018 to create in-browser notebooks for scientific exploration and visualization in a similar vein to Jupyter:

Iodide is no longer actively maintained. Pyodide can however be used in other interactive client-side notebook environments (e.g., Starboard, Basthon, JupyterLite).

The full release note and announcement are available online and contain plenty of additional illustrations and explanations. Pyodide is now an independent and community-driven open-source project distributed under the Mozilla Public License Version 2.0.

Read the original here:

Pyodide Brings Python and Its Scientific Stack to the Browser with WebAssembly - InfoQ.com

The Indian Institute of Technology (IIT), Kanpur, has rolled out two new courses in the field of statistics and data science. The academic senate and board of the institute have approved new 4-year BS and 5-year BS-MS programmes in Statistics and Data Science from the coming academic year 2021-22. The admissions to the programmes will be through JEE (Advanced) score.

Data Science, Artificial Intelligence/ML are playing increasingly important roles in finding solutions to diverse real-world problems. Research in data science in the next 20 years is expected to focus on developing mathematically rooted models that can be implemented, Abhay Karandikar, director, IIT Kanpur said.

The synergy of theory and applications requires training specific to a unique set of skills, and the BS and BS-MS programmes in Statistics and Data Science have been launched in response to this need for effective implementation in real-life situations, Karandikar added.

Read | IIT Delhi launches new centre for research in optics and photonics

In collaboration with the proposed School of Medical Research and Technology (SMRT), students would also work on health-related data and digital health to give a much-needed boost to research and analytics in this important emerging area.

Students undergoing this programme will be exposed to various types of structured and unstructured data with applications and will not only be well suited to build flourishing careers in industry and the new entrepreneurial India but also to pursue higher studies, Karandiar said.

Read the original:

IIT Kanpur to offer two new courses in Statistics and Data Science from 2021-22 session - The Indian Express