- Open Access
Analysis of biomedical data with multilevel glyphs
© Müller et al.; licensee BioMed Central Ltd. 2014
- Published: 16 May 2014
This paper presents multilevel data glyphs optimized for the interactive knowledge discovery and visualization of large biomedical data sets. Data glyphs are three- dimensional objects defined by multiple levels of geometric descriptions (levels of detail) combined with a mapping of data attributes to graphical elements and methods, which specify their spatial position.
In the data mapping phase, which is done by a biomedical expert, meta information about the data attributes (scale, number of distinct values) are compared with the visual capabilities of the graphical elements in order to give a feedback to the user about the correctness of the variable mapping. The spatial arrangement of glyphs is done in a dimetric view, which leads to high data density, a simplified 3D navigation and avoids perspective distortion.
We show the usage of data glyphs in the disease analyser a visual analytics application for personalized medicine and provide an outlook to a biomedical web visualization scenario.
Data glyphs can be successfully applied in the disease analyser for the analysis of big medical data sets. Especially the automatic validation of the data mapping, selection of subgroups within histograms and the visual comparison of the value distributions were seen by experts as an important functionality.
- Interactive Knowledge Discovery
- Semantic Zoom
Professionals in the biomedical domain are confronted with increasing masses of data, which require efficient and user-friendly solutions and the development of methods to assist them in knowledge discovery to identify, extract, visualize and understand useful information from these large amounts of data . The trend towards personalized medicine has resulted in a mass of clinical, laboratory and genome-scale data and moreover, most data models are characterized by complexity, which makes manual analysis very time-consuming and frequently practically impossible . The major challenge is: How can an expert find knowledge in these terabytes of complex data? For example, to successfully search for novel hypotheses in large datasets, we must look for unexpected patterns and interpret evidence in ways that frame new questions and suggest further explorations. Consequently, methods from Knowledge Discovery and Visual Analytics methods may help us to
Overview large data sets as the human visual sense is optimized for parallel processing
Connect the global view with detail information
Provide different contextual views (e.g. expert versus common user)
Deal with inhomogeneous data sets and broad range of data quality.
As one solution to these goals, we developed a set of validated glyphs for interactive exploration of biomedical data sets. With the ability to work with different level of details, to arrange and order the glyphs in space and to synchronise different visualizations through coordinated multiple views (CMV) , an expert can in the truest sense of the word, travel through his data space.
Jacques Bertin's book Sémiologie graphique, published in 1967 (English translation 1987 by J. Berg), provides the foundation for the analysis of visual elements to display qualitative or quantitative data . Bertin's practical experience as a cartographer led him to the question how to find rules to build proper graphics. His study of signs together with their "grammatical" rules is based on a clear and logical symbol scheme in which symbols can be varied referring to visual variables. Visual variables include size of elements, their shape, orientation, brightness color, texture and position. Bertin called these attributes also retinal variables, because they describe the quality characteristics of the human perception, in contrast to a technical description of a graphical element. Actually, this leads to semiotics - and we view informatics as semiotics engineering , because it is interesting to observe that the three main goals of informatics (correctness of algorithms, efficiency of programs, and usability of software systems) turn out to be nicely related to the three semiotic dimensions : 1) Correctness is a matter of syntax to be answered by considering formal aspects only ; 2) Efficiency is a matter of semantics related to the object world ; and 3) Usability, taking interest and motivation of the end user into account ; being our basic assumptions for the following details:
A visual variable is characterized according to Bertin by the kind of scale (nominal ordinal) and the length of the visual variable. The length of a variable is the number of distinguishable values that can be perceived by a viewer (for example how many shades of grey or different hue values can be differentiated) Choosing different visual variables for representing the same data variable greatly influence the perception and understanding of the glyph. It is therefore important to know and appropriately map data variables to visual variables in the design of a glyph.
Our approach will make use of visual variables to describe the perceptual properties of a glyph. Ropinski & Preim (2008) and Ropinski, Oeltze & Preim (2011) ,  describe glyph-based visualization techniques in medical visualizations and give a glyph taxonomy together with guidelines for the usage of glyphs. Ward (2002)  describes a taxonomy of glyph placement strategies, were he distinguishes between data-driven and structure-driven approaches. He also describes strategies to avoid overlapping problems and proposes a spacefilling layout for structured data.
A very specific type of glyphs was introduced by Chernoff (1973): the so-called Chernoff faces . Chernoff faces are 2D glyphs, which employ human's ability to recognize faces and small changes in facial characteristics. However the effectiveness of this form of visualization is still being debated in the scientific community , .
Kraus & Ertl  present in a more technical approach a system for glyph generation (with minimal user interaction) which has been used in a visualization tool in the automotive industry.
An overview about the state of the art in the visualization of multi-variate data is given by Peng & Laramee (2009)  as well as Bürger & Hauser (2007), where they discuss how different techniques take effect at specific stages of the visualization pipeline and how they apply to multi- variate data sets being composed of scalars, vectors, and tensors. Moreover they provide a categorization of these techniques in the aim for a better overview of related approaches , with an update published 2009 . Visual data exploration methods on large data sets were described by several authors, and particularly Keim (2001) , Hege et al. (2001) , Fayyad, Wierse & Grinstein (2002), , Fekete & Plaisant (2002) , and Santos & Brodlie (2004)  provide a good introduction to this topic. A recent state-of-the-art report on glyph based visualization and a good overview on theoretic frameworks, e.g. on the semiotic system of Bertin, was given by Borgo et. al. (2013) .
An interesting application of glyphs for a visual analytics approach for understanding biclustering results from microarray data has been presented by Santamaria, Theron & Quintales (2008),  and another one by Gehlenborg & Brazma (2009),  and Helt et al (2009),  and a recent work by Konwar et al (2013), .
The closest work to use glyphs with an adaptive layout is the work of Legg et al. (2012)  in the application domain of sport analysis. Here the data space is event based, and the adaptive layout strategy is focused on overlapping events with so called "macro glyphs", which combine several glyphs into one. In the "macro glyph" approach only scaling and no level of detail (LoD) suitable for different screen spaces are applied. In the evaluation phase expert interviews at the work environment level based on methods described by Tory & Möller (2004)  and Plaisant (2004)  were done.
Our previous work [34, 35] in biomedical visualization resulted in an upper bound of 16 attributes for the highest level of detail. This number is given be the attribute set in a pathological finding, which is composed of patient information (age, sex, year of birth, year of death, cause of death, disease free survival), the pathological finding (organ, size of the tumor, lymph nodes staging, metastasis staging, grading, receptor state ) and surgery attributes (origin of the sample, year of surgery, doctor, type of sample). In order to unveil hidden relations by the recognition of unexpected patterns, as many variables as possible should be integrated within the rendering of one glyph. 2D glyph designs are usually limited to up to 5 data variables, therefore we chose the approach to model data glyphs as 3D objects. This results on the one hand a high information density but on other hand we face the problems of occlusion, perspective distortion and complex navigation and orientation in 3D space. Usability tests with very first prototypes have indicated that glyphs placement in 3D space using a perspective projection and the possibility to freely move within this space was overly burdensome for almost all users, especially for medical experts. To avoid the problems described above, we restricted the 3D space to 2.5D or to a ¾ perspective view by applying dimetric (near isometric) projection grid, well known from technical illustrations and from some very successful simulation games of the 1990s (e.g Civilization ) In a diametric projection grid data glyphs do not change size as they are moved, so no re-rendering of a glyph is necessary to simulate a ¾ perspective view. With a dimetric projection grids also specific performance optimization strategies, e.g. bitmap caching and selection highlighting can be easily applied.
Level of detail
As we want visualize several millions data elements in the smallest level of detail, the screen size of a glyph can be as small as one pixel. Therefore only the visual variable "value" (from light to dark) or "color" (changes in hue at a given value) can be the starting point. Note: If the maximal number of elements to be visualized is in the range of several 10.000 elements, we can also choose the visual variable shape as starting point. To achieve well-graduated levels of details and visually smooth transition between leves we rely at the principle that the dominant visual variable of level n is also the strongest visual variable in level n+1.
Visual Variables of the Cubic Glyph
height main cube
color west-element base
color cap east-element
height east -element
color east -element base
color cap east -element
color south-element base
color cap south-element
According to the taxonomy given by Ward  we support:
User driven placement, in which case the user determines the position of a glyph through interaction tasks (selection, filtering, movement, grouping)
Data driven placement, in which case data values are used to specify the location of the glyph. Our placement strategy supports value discretization and jittering strategies for the placement in an dimetric projection grid,
Structure driven placement, in which case relationship between data points determines the location of a glyph. We support structure directly derivable from the data values, e.g. grouping glyph representing cancer cases by year of surgery, sex and cancer staging, and glyph placements determined by interactive ant clustering algorithm.
A data glyph can be configured through the mapping of data variables to the parameters of its geometric primitives. This is on the one hand a very powerful tool, as the user can map any data attribute to any geometric parameter, and even change the mapping on the fly, on the other hand its also crucial, because the great flexibility could easily lead to faulty mappings (e.g. mapping a nominal variable to the position of a geometrical primitive) and in succession to misinterpretations of the visualizations results. In order to avoid those mismatches we provide an automatic validation of the variable mapping.
In the automatic validation, we compare meta information about data variables - scale of measurement (discrete, continuous, categorical, ordinal, interval, nominal) and the number of distinct values - to the visual capabilities of the glyph elements. The verification is done according to the following rules:
The shape of a geometric primitive is purely nominal and should therefore never be mapped to ordinal data values. However we can recognize a almost infinite variety of shapes (the shape variable is "very long").
The perceptual variable color (hue) is a nominal variable, even though the wavelength of light assigns an ordering to colors, the human perceptual system takes no notice of it. There is some cultural ordering imposed on hue (red is "hotter" than blue), but it is weak because not all hues are related. A non-color deficient person can distinguish between seven and ten million different colors. However, color is a deeply subjective attribute, and therefore not more than 10 to 20 carefully chosen color values should be used in color mapping. A great tool for carefully designed colormaps, which e.g. provides "colorblind safe" suggestions, can be found at http://colorbrewer2.org
Value (the brightness of an element) and the texture (with respect to the grain size of the texture) are ordered and can be mapped to an ordinal scale. Value and texture are short variables, i.e. roughly 10 values can be distinguished in an effective way.
The position of a glyph can be mapped to ordinal values, and is a very fine-grained (long) variable. The size of a geometric primitive, or even of the whole glyph element can also be mapped to ordinal values, but it is "shorter" than the position variable.
Finally the orientation of a geometric primitive can be mapped to an ordinal data value, but this is a very short viusal variable, i.e. only very few different orientations can be perceived.
We use multilevel data glyphs in the disease analyser, a visual analytic application for the interactive exploration of a database containing approximately 1,4 million cancer cases. Each record describes a comprehensive diagnosis of a cancerous (malignant) tumor case. The most used variables are patient age and sex, the ICDN classification, the TNM staging, grading receptor states and information about the time under risk, disease free survival and overall survival together with surgery information.
Green: All data scales fits to the scale of corresponding visuals variable the length of all visual variables is equal/greater then the corresponding distinct data values.
Yellow: All data scales fits to the scale of visuals variables and the length of some visual variable is smaller then the number of corresponding distinct data values.
Red: There is a mismatch (minimal one) attribute scale and the scale of the corresponding visual variable.
The utilization of multilevel data glyphs in the disease analyser was a valuable source for the development of our glyph design criteria. In the design process we faced the following challenges:
Occlusion: 3D glyphs provide on the one hand high data density, but on the other hand face the problem of occlusion. To minimize the occlusion effect we put the main visual variable on top of the geometry (especially in the iconic view) and limit the height of the data glyph. Perspective distortions are avoided by the use a parallel projection (2½D view of an object with forced depth). We use either a dimetric projection or a cavalier or military projection when the glyphs should be seen from a higher point of view.
Secondary colors: Multilevel glyphs consist of complex geometry, where each geometric primitive can be colored independently. This may result in undesirable secondary (mixed) colors. To avoid this effect a good glyph design provides a clear gradation of visual variables, especially for color perception. Such a gradation can be achieved through well defined increments of the graphic primitives size and a restricted color mapping for individual graphical primitives. In some special cases secondary colors could be used intentionally, e.g. to visualize the coincidence of two values in a large data set.
Grid patterns: When data glyphs are arranged in a dense grid unwanted patterns can occur. To avoid this, a good glyph design is based on a symmetrical skeletal structure. Especially in the iconic view it is crucial to model borders of the glyph, in order to provide a good visual differentiation. In the simplest case a border can be realized through a plinth as a neutral base element.
During beta testing the disease analyser was used by 12 experts working in the field of bioinformatics, computational biology and medical research. The first group had a focus on data acquisition, automatic classification of medical records and data quality issues. The focus of the second group was on data analysis, e.g. the development of the health care system, and hypothesis generation. The following observations and statements describe their experience and provide valuable input for further developments:
The disease analyser is very well suited to find outliers and "white spaces" in the source data.
Snapshot and bookmarking functionality is missing.
The selection of subgroups within the histograms and the visual comparison of the value distributions were very much appreciated.
In research tasks, the disease analyser was used to compare two to four subgroups.
Manual arrangement and sorting of cases was used often.
The fast availability of the full diagnosis text for the selected data glyph is an important feature.
When a hypothesis is generated there should be a report module to (statistically) compare the involved subgroups and to print out a report.
We developed multilevel data glyphs for the visualization of large medical data sets. The data glyphs provide
three levels of detail (semantic zoom) suitable for a different screen space, and a
validation of the data variable mapping.
We used multilevel data glyphs in the disease analyser, a visual analytic application for quality control and exploration of a comprehensive collection of cancer disease records. Three concrete glyph designs and design rules resulted out of the hands-on- experience.
We plan to integrate the proposed data glyphs as a visual front end to the biobank of the Medical University Graz and for quality assurance tasks of data record related to cancer samples and to apply the visualization method for strategic planning and trend analysis in the medical domain. In the undertaking we will use a lightweight (webGL) version of data glyphs, which can be used as visualization components in a webpage connected to a local datagrid or through a web service to a central medical database.
There are a lot of studies to compare of 2D versus 3D visualization techniques for the visualizations of spatial related data, e.g. medical renderings or geographic data. However there is now systematic evaluation known to the authors comparing 2D glyphs to 3D and 2½D (isometric) techniques for abstract information. For abstract information no inherent mapping of the data either to the 3D shape of a glyph nor the spatial position is given, which would be a natural mental model for users of the visualization results. Lie et al  have discussed design and realization aspects (occlusion, depth perception and visual cluttering) of glyph based 3D-data visualization with a focus on glyph placement. Their work is a good starting point for a systematic evaluation of the shape/placement of 2½D glyphs providing high data density versus 2D shapes, which are less challenging for the user perception.
A second open research question is how to build and evaluate smooth transitions between different levels of glyph abstraction. In the current work the glyph rendering method was changed due to the glyph size in the screen space. The configuration of "switching points" was done with a heuristic approach, and carefully (manual) designed glyph geometry resulted in a smooth visual transition. However a systematic study and description of the methodology of glyph transitions (fusion of semantic and graphical zoom) has still to be done.
This work was funded by the FIT-IT programme (813 398) and by the Austrian Fonds zur Förderung der wissenschaftlichen Forschung (FWF, L427-N15). Medical data were provided in the context of the Austrian Genome Programme GEN-AU and the CRIP project. Our thanks are due to all partners, for their contributions, critical reviews and various discussions. The work has been approved by the Ethical Committee of the Medical University of Graz.
Publication for this article has been funded by the Christian Doppler laboratory for biospecimen research and biobanking technologies.
This article has been published as part of BMC Bioinformatics Volume 15 Supplement 6, 2014: Knowledge Discovery and Interactive Data Mining in Bioinformatics. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcbioinformatics/supplements/15/S6.
- Holzinger A, Zupan M: KNODWAT: A scientific framework application for testing knowledge discovery methods for the biomedical domain. BMC Bioinformatics. 2013, 14 (1): 191-10.1186/1471-2105-14-191.PubMed CentralView ArticlePubMedGoogle Scholar
- Holzinger A: Human-Computer Interaction & Knowledge Discovery (HCI-KDD): What is the benefit of bringing those two fields to work together?. Multidisciplinary Research and Practice for Information Systems, Springer Lecture Notes in Computer Science LNCS 8127. Edited by: Alfredo Cuzzocrea CK, Dimitris E. Simos, Edgar Weippl, Lida Xu. 2013, Heidelberg, Berlin, New York: Springer, 319-328.Google Scholar
- Turkay C: Hypothesis Generation by Interactive Visual Exploration of Heterogeneous Medical Data. Human-Computer Interaction and Knowledge Discovery in Complex, Unstructured, Big Data. 2013, Springer Berlin Heidelberg, 1-12.View ArticleGoogle Scholar
- Baldonado M, Woodruff A, Kuchinsky A: Guidelines for Using Multiple Views in Information Visualization. in Proc of Advanced Visual Interfaces (AVI 2000). 2000, ACM Press, 110-119.Google Scholar
- Bertin J, Barbut M: Semiology of Graphics Diagrams Networks Maps. Edited by: J. Berg. 1983, University of Wisconsin Press, (French edn. 1967)Google Scholar
- Holzinger A, Searle G, Auinger A, Ziefle M: Informatics as Semiotics Engineering: Lessons Learned from Design, Development and Evaluation of Ambient Assisted Living Applications for Elderly People. Universal Access in Human-Computer Interaction Context Diversity, Lecture Notes in Computer Science, LNCS 6767. Edited by: Stephanidis C. 2011, Berlin, Heidelberg: Springer, 183-192.Google Scholar
- Andersen PB: What Semiotics can and cannot do for HCI. Knowledge- Based Systems. 2001, 14 (8): 419-424. 10.1016/S0950-7051(01)00134-4.View ArticleGoogle Scholar
- Hoare AR: Proof of correctness of data representations. Acta Informatica. 1972, 1 (4): 271-281. 10.1007/BF00289507.View ArticleGoogle Scholar
- Nake F, Grabowski S: Human-Computer Interaction viewed as Pseudo- Communication. Knowledge-Based Systems. 2001, 14 (8): 441-447. 10.1016/S0950-7051(01)00140-X.View ArticleGoogle Scholar
- Holzinger A: Usability engineering methods for software developers. Communications of the ACM. 2005, 48 (1): 71-74. 10.1145/1039539.1039541.View ArticleGoogle Scholar
- Ropinski T, Preim B: Taxonomy and Usage Guidelines for Glyph-based Medical Visualization. SimVis - Simulation and Visualization:. 2008, SCS Publishing House, 121-138. ; MagdeburgGoogle Scholar
- Ropinski T, Oeltze S, Preim B: Survey of glyph-based visualization techniques for spatial multivariate medical data. Computer & Graphics. 2011, 35 (2): 392-401. 10.1016/j.cag.2011.01.011.View ArticleGoogle Scholar
- Ward MO: A taxonomy of glyph placement strategies for multidimensional data visualization. Information Visualization. 2002, 1 (3-4): 194-210. 10.1057/palgrave.ivs.9500025.View ArticleGoogle Scholar
- Chernoff H: Use of Faces to Represent Points in K-Dimensional Space Graphically. 1973, J Am Stat Assoc, 68 (342): 361-368.Google Scholar
- Morris CJ, Ebert DS, Rheingans PL: Experimental analysis of the effectiveness of features in Chernoff faces. 28th AIPR Workshop: 3D Visualization for Data Exploration and Decision Making:. 2000, International Society for Optics and Photonics: 12-17Google Scholar
- Lee MD, Reilly RE, Butavicius ME: An empirical evaluation of Chernoff faces, star glyphs, and spatial visualizations for binary data. Proceedings of the Asia-Pacific symposium on Information visualisation Volume 24. 2003, Australian Computer Society, Inc, 1-10.Google Scholar
- Kraus M, Ertl T: Interactive Data Exploration with Customized Glyphs. International Conferences in Central Europe on Computer Graphics, Visualization and Computer Vision (WSCG 2001): 2001; Pilzen (Czech Republic). 20-23.Google Scholar
- Peng Z, Laramee S: Higher Dimensional Vector Field Visualization. 2009, A Survey in Theory and Practice of Computer Graphics (TPCG '09), 149-163.Google Scholar
- Bürger R, Hauser H: Visualization of multi-variate scientific data. 2007, Proceedings of EuroGraphics, 117-134.Google Scholar
- Fuchs R, Hauser H: Visualization of Multi-Variate Scientific Data. Computer Graphics Forum:. 2009, Wiley Online Library, 1670-1690.Google Scholar
- Keim DA: Visual exploration of large data sets. Communications of the ACM. 2001, 44 (8): 38-44. 10.1145/381641.381656.View ArticleGoogle Scholar
- Hege H-C, Hutanu A, Kähler R, Merzky A, Radke T, Seidel E, Ullmer B: Progressive retrieval and hierarchical visualization of large remote data. Scalable Computing: Practice and Experience. 2001, 6 (3): 60-72.Google Scholar
- Fayyad UM, Wierse A, Grinstein GG: Information visualization in data mining and knowledge discovery. 2002, Morgan KaufmannGoogle Scholar
- Fekete J-D, Plaisant C: Interactive information visualization of a million items. Information Visualization. 2002, IEEE, 117-124. INFOVIS 2002 IEEE Symposium on: 2002Google Scholar
- Dos Santos S, Brodlie K: Gaining understanding of multivariate and multidimensional data through visualization. Computers & Graphics. 2004, 28 (3): 311-325. 10.1016/j.cag.2004.03.013.View ArticleGoogle Scholar
- Borgo R, Kehrer J, Chung DH, Maguire E, Laramee RS, Hauser H, Chen M: Glyph-based Visualization: Foundations, Design Guidelines, Techniques and Applications. Eurographics 2013-State of the Art Reports. 2012, The Eurographics Association, 39-63.Google Scholar
- Santamaría R, Therón R, Quintales L: A visual analytics approach for understanding biclustering results from microarray data. BMC Bioinformatics. 2008, 9 (1): 247-10.1186/1471-2105-9-247.PubMed CentralView ArticlePubMedGoogle Scholar
- Gehlenborg N, Brazma A: Visualization of large microarray experiments with space maps. BMC Bioinformatics. 2009, 10 (Suppl 13): O7-10.1186/1471-2105-10-S13-O7.PubMed CentralView ArticleGoogle Scholar
- Helt G, Nicol J, Erwin E, Blossom E, Blanchard S, Chervitz S, Harmon C, Loraine A: Genoviz Software Development Kit: Java tool kit for building genomics visualization applications. BMC Bioinformatics. 2009, 10 (1): 266-10.1186/1471-2105-10-266.PubMed CentralView ArticlePubMedGoogle Scholar
- Konwar KM, Hanson NW, Pagé AP, Hallam SJ: MetaPathways: a modular pipeline for constructing pathway/genome databases from environmental sequence information. BMC Bioinformatics. 2013, 14 (1): 202-10.1186/1471-2105-14-202.PubMed CentralView ArticlePubMedGoogle Scholar
- Legg PA: MatchPad: Interactive Glyph Based Visualization for Real Time Sports Performance Analysis. Computer Graphics Forum. 2012, Blackwell Publishing Ltd, 31 (3pt4): 1255-1264. 10.1111/j.1467-8659.2012.03118.x.Google Scholar
- Tory M, Moller T: Human factors in visualization research. Visualization and Computer Graphics, IEEE Transactions on. 2004, 10 (1): 72-84. 10.1109/TVCG.2004.1260759.View ArticleGoogle Scholar
- Plaisant C: The challenge of information visualization evaluation. Proceedings of the working conference on Advanced visual interfaces:. 2004, ACM, 109-116.View ArticleGoogle Scholar
- Müller H, Zatloukal K, Streit M, Schmalstieg D: Interactive Exploration of Medical Data Sets. Proceedings of the Conference on BioMedical Visualisation. 2008, London, UK, 29-35.Google Scholar
- Müller H, Reihs R, Sauer S, Zatloukal K, Streit M, Lex A, Schlegl B, Schmalstieg D: Connecting Genes with Diseases. Proceedings of the13th International Conference on Information Visualisation. 2009, Barcelona, 323-330.Google Scholar
- Greene F: AJCC cancer staging handbook. American Joint Committee on Cancer;. 2002, New York : SpringerGoogle Scholar
- Harrower M, Brewer CA: ColorBrewer. org: an online tool for selecting colour schemes for maps. Cartographic Journal. 2003, 40 (1): 27-37. 10.1179/000870403235002042.View ArticleGoogle Scholar
- Lie AE, Kehrer K, Hauser H: Critical design and realization aspects of glyph-based 3D data visualization. Proceedings of the 25th Spring Conference on Computer Graphics. 2009, ACM, 27-34.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.