Visual Analytics of Image-Centric Cohort Studies in Epidemiology

01/15/2015 ∙ by Bernhard Preim, et al. ∙ Otto-von-Guericke-University Magdeburg 0

Epidemiology characterizes the influence of causes to disease and health conditions of defined populations. Cohort studies are population-based studies involving usually large numbers of randomly selected individuals and comprising numerous attributes, ranging from self-reported interview data to results from various medical examinations, e.g., blood and urine samples. Since recently, medical imaging has been used as an additional instrument to assess risk factors and potential prognostic information. In this chapter, we discuss such studies and how the evaluation may benefit from visual analytics. Cluster analysis to define groups, reliable image analysis of organs in medical imaging data and shape space exploration to characterize anatomical shapes are among the visual analytics tools that may enable epidemiologists to fully exploit the potential of their huge and complex data. To gain acceptance, visual analytics tools need to complement more classical epidemiologic tools, primarily hypothesis-driven statistical analysis.



There are no comments yet.


page 18

page 20

page 24

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Epidemiology is a scientific discipline that provides reliable knowledge for clinical medicine focusing on prevention, diagnosis and treatment of diseases Fletcher2011 . Research in epidemiology aims at characterizing risk factors for the outbreak of diseases and at evaluating the efficiency of certain treatment strategies, e.g., to compare a new treatment with an established gold standard. This research is strongly hypothesis-driven and statistical analysis is the major tool for epidemiologists so far. Correlations between genetic factors, environmental factors, life style-related parameters, age and diseases are analyzed. The data are acquired by a mixture of interviews (self-reported data, e.g., about nutrition and previous infections) and clinical examinations, such as measurement of blood pressure. Statistical correlations, even if they are strong, may be misleading because they do not represent causal relations. As an example, the slightly reduced risk of heart infarct and cardiac mortality for elderly people reporting to drink one glass of wine every evening (compared to people drinking no alcohol at all) may be due to the involved low level of alcohol but may also be a consequence of a very regular and stress-free lifestyle Fletcher2011 . When something happened, before an event, it is an indicator for a causal relationship. However, care is necessary, since many things happen in the life of an individual before, e.g., a heart attack, but do not cause it.

Thus, statistical correlations are the starting point for investigating why certain factors increase the risk of getting diseases. Epidemiology is not a purely academic endeavor but has huge consequences for establishing and evaluating preventive measures even outside of medicine. The protection of people from passive smoking, recommendations for various vaccinations and the introduction of early cancer detection strategies, e.g., mammography screening, are all based on large-scale epidemiological studies. Also the official guidelines for the treatment of widespread diseases, such as diabetes, are based on evidence from epidemiological studies Fletcher2011 . While this all may sound obvious, it is a rather recent development. Evidence-based medicine often still has to “fight” against recommendations of a few opinion leaders arguing based on their personal experience only.

The analysis techniques used so far are limited to investigating hypotheses based on known or suspected relations, e.g. hypotheses related to observations or previous publications. The available tools support the analysis of a few dimensions, but not of the hundreds of attributes acquired per individual in a cohort study. Both typical visualization techniques as well as analysis techniques, e.g., support vector machines, do not scale well for hundreds of attributes

Turkay:2013 . While we are not able to describe solutions for these challenging problems, we give a survey on recent approaches aiming also at hypothesis generation.

Organization. This chapter is organized as follows. In Sect. 2 we describe important concepts and terms of epidemiology including observations from epidemiologic workflows. This discussion is restricted to those terms that are crucial for communicating with epidemiologists, understanding requirements and for designing solutions that fit in their process. In Sect. 3, we discuss how (general) information visualization and data analysis techniques may be used for epidemiologic data. Section 4 describes the analysis of image data from cohort studies and how this analysis is combined with the exploration of non-image attribute data. This section represents the core of the chapter and employs a case study where MRI data of the lumbar spine are analyzed along with attributes characterizing life-style, working habits, and back pain history.

2 Background in Epidemiology

Population-based studies. Epidemiological studies are based on a sample of the population. The reliability of the results obviously depends on the size of that sample but also strongly on the selection criteria. Often, data from patients treated in one hospital are analyzed. While this may be a large number of patients, the selection may be heavily biased, e.g., since the hospital is highly specialized and diseases are often more severe or in a later stage compared to the general population.

Population-based studies, where representative portions of a population (without known diseases) are examined, have the potential to yield highly reliable results. The source population may be from a city, a region or a country. Individuals are randomly selected, e.g., approaching data bases of population registries. The higher the percentage of people who accept the invitation and actually take part in the study, the more reliable the results are.

In this chapter, we focus on longitudinal population-based studies. The sheer amount and diversity in terms of type of data makes it difficult to fully identify and analyze interesting relations. We will show that information visualization and visual analytics techniques may provide substantial support that complements the statistical tools with their rather simple statistical graphics. Most epidemiological studies were restricted to nominal (often called categorical) and scalar data, e.g., related to alcohol consumption, and body mass index as one measure of obesity.

Image-centric epidemiological studies. More recently, for example, in the Rotterdam study Hofman2009 , also non-invasive imaging data, primarily ultrasound and MRI data, are employed. Petersen and colleagues Peterson2013

report on six studies involving cardiac MRI from at least 1000 individuals in population-based studies. These high-dimensional data enable to answer analysis questions, e.g., how does the shape of the spine changes as a consequence of age, life style and diseases? We focus on such

image-centric epidemiological studies.

Epidemiology and public health. There are different branches of epidemiology. One branch deals with predictions to inform public health activities. These include measures in case of an epidemic – an acute public health problem, mostly related to infectious diseases. The recent article ”computational epidemiology” Marathe2013

was focussed on this branch of epidemiology. Another branch of epidemiology aims at long-term studies and at findings primarily essential for prevention. Image-centric cohort studies, the focus of this article, belong to this second branch. The target user group consists of epidemiologists who can be expected to have a high level of expertise in statistics. Thus, their findings involve statistical significance, confidence intervals and other measures of statistical power.

Healthy aging and pathologic changes. An essential problem in the daily clinical routine is the discrimination between healthy age-related modifications (that may not be reversed by treatment) and early stage diseases (that may benefit from immediate treatment). As a consequence, elderly people are often not adequately treated. As a general goal for epidemiological studies, better and more reliable markers for early stage diseases are searched for. The cardiovascular branch of the Rotterdam study, for example, aims at an understanding of atherosclerosis, coronary heart disease and “cardiovascular conditions at older age” Hofman2009 .

Modern epidemiology. Epidemiology faces new challenges due to the rapid progress, e.g., in genetics and sequencing technology as well as medical imaging. Acquisition of health data thus becomes cheaper and more precise. In cohort studies, as much potentially relevant data as possible are acquired as a basis for an as broad as possible spectrum of analysis questions. This includes blood, urine and tissue samples, information about environmental conditions and the social milieu.

Visual analytics for modern epidemiology. In the past, epidemiology primarily dealt with hypotheses aiming to prove them, e.g., the efficiency of early cancer detection programs in terms of mortality and long term survival Fletcher2011 . Since recently, more and more data mining is performed to identify correlations. Results of such analyses, however, need to be very carefully interpreted. If thousands of potential correlations are analyzed automatically, just by chance some of them will reach a high level of statistical significance.

An essential support for epidemiology research is to define relevant subgroups. To perform separate analyses for women and men as well as for different age groups is a common practice in epidemiology. However, relevant subgroups may be defined by a non-obvious combination of several attributes that may be detected by a combination of cluster analysis and appropriate visualization.

Since the information space is growing with each examination cycle, Pearce and Merletti Pearce2006 pointed out in 2006 that methods are needed which can cope with this complexity and enable the analysis of underlying causes of a certain disease. Visual analytics (VA) methods can support epidemiological data assessment in different ways, e.g. by defining subgroups based on a multitude of attributes that exhibit a certain characteristic. For the analysis of scalar and categorical data, established information visualization techniques combined with clustering and dimension reduction are a good starting point, but need to be tightly integrated with statistic tools epidemiologists that are more familiar with. For image-centric studies, however, new visualization, (image) analysis and interaction techniques are needed.

In the following, we define essential terms in epidemiology and give an overview on cohort studies that employ medical image data as an essential element. Finally, we describe how image data, derived information and other data complement each other to identify and characterize risks.

2.1 Important Terms

Prevalence and incidence. Epidemiology investigates how often certain diseases or clinical events, such as a cerebral stroke or sudden heart death, occur in the population. Two terms are important to characterize this frequency. The prevalence indicates the portion of people suffering from a disease at a given point in time. The incidence

represents how many people suffer from a disease or event in a certain interval, usually one year. High prevalence is usually associated with high economic costs. Population-based studies focus on diseases with a high prevalence, such as diabetes, coronary heart disease or neurodegenerative diseases. Even these diseases do not occur frequently in a random population including many younger people (where the prevalence of these diseases is low). A rare disease, such as amyotrophic lateral sclerosis, may have a prevalence of 5 from 100,000. Thus, even in a large population-based study probably no individual suffers from this disease.

Absolute and relative risks. Another essential epidemiological term is the risk for a clinical event, such as outbreak of a certain disease, severity (stage) or death. As an example, a study related to cardiac risk may investigate angina pectoris, myocardial infarction, atrial fibrillation depending on attributes such as age and sex. The absolute risk characterizes the likelihood of getting a disease in life time. The absolute risk for a woman to develop breast cancer in the Western world is particularly high for women aged 50-60 (2.6%) and 60-70 (3.7%). Therefore, for these age groups, mammography screening – aiming at early detection and thus optimal treatment – was introduced.

The relative risk (RR) characterizes the increased risk if an individual is exposed to a certain risk factor, e.g., smoking, excessive weight, or alcohol abuse. It is based on a comparison with a control group not exposed to that risk factor. A value of represents a factor that protects, e.g., moderate physical activity. Exciting observations are often the combined effects of several parameters. A certain factor may be protective for some people (younger, slim women) and is involved with an increased risk for others. The combined risk may be significantly smaller or larger than could be expected from individual factors.

Moreover, relationships are often distinctly non-linear or even non-monotonic. Dose-response relationships are often non-linear. RR increases slowly (almost no effect for a small dose) and increases much faster for higher levels of a dose, e.g., exposure to toxicity. A typical non-monotonic relation is U-shaped, that is both very low and very high instances of an attribute involve an increased risk, whereas values in between are associated with a reduced risk. Examples are weight (both very low and very high weight are associated with an increased risk for mortality) and sleeping time (both very short and very long sleepers have an increased risk for developing psychiatric disorders Hofman2009

). Such relations cannot be characterized by a global RR value. Instead, tools are necessary that support the hypothesis of a U-shaped relation by estimating their parameters with some kind of best-fit algorithm.

2.2 Image-Centric Cohort Studies

Image data in epidemiology. The acquisition of image data is determined by the available time, by financial resources, by the epidemiological importance and by ethic considerations. Epidemiological studies require approval by a local ethics committee. As a consequence, healthy individuals in a cohort study should not be exhibited to a risk associated to the examinations carried out. Thus, MRI should be preferred over X-ray or CT imaging for its non-radiation nature. Petersen and colleagues Peterson2013 explain why cardiac CT is less feasible in a cohort study and even MR is only used without a contrast agent in their study due to ethical reasons. MRI data and ultrasound data are the prevailing modalities in both the SHIP as well as the Rotterdam study. Unfortunately, MRI and ultrasound data do not exhibit standardized intensity values (in contrast to CT data). Moreover, MRI and ultrasound data suffer from inhomogeneities and various artifacts. Thus, they are more difficult to interpret for humans and more difficult to analyze with computational means. These data are used to measure, e.g., the thickness of vessel walls, the abdominal aorta diameter and plaque vulnerability in the coronary vessels Hofman2009

. The intensive use of MRI in epidemiological research also explains to some extent which questions are analyzed: MRI is the best modality for the analysis of brain structures and thus serves to explore early signs of Parkinson’s, Alzheimer’s and other neurodegenerative diseases. Epidemiological research aims at identifying such brain pathologies in a pre-symptomatic stage. Among the sources for such investigations are MR Diffusion Tensor Imaging data that enable an assessment of white matter integrity

Hofman2009 .

The selection of imaging parameters is always a trade-off between conflicting goals related to quality, e.g., image resolution, signal-to-noise ratio, patient comfort, e.g., examination time and associated costs. As a consequence, to shorten overall examination times in cohort study examinations, not the highest possible quality is available, i.e., a slice distance of 4 mm is more typical than 1 mm. A great advantage of MRI is that this method is very flexible and enables to display different structures in different sequences, such as T1-, T2- and proton density-weighted imaging. MRI data in cohort studies often comprise more than ten different sequences.

Standardization in image acquisition. Due to the rapid progress in medical imaging, sequences, protocols and even (MR) scanners are frequently updated in clinical routine (similar to the update frequency on a computer). These updates would severely hamper the comparison of imaging results and thus the assessment of natural changes and disease outbreak. Thus, differences in acquisition parameters are essential confounding variables

. Therefore, for one cohort and examination cycle that may last up to several years, no updates are allowed. Moreover, all involved physicians and radiology technicians are carefully instructed to use the same standardized imaging parameters. This point is even more important for longitudinal studies with repeated imaging examinations. Even if MR scanners and protocols are not updated, the life cycle of MR coils leads to changes of image quality that need to be monitored and compensated.

2.3 Examples for Image-Centric Cohort Study Data

In the following, we describe selected comprehensive and on-going longitudinal cohort studies. Both use a number of (epidemiologic) instruments that are innovative in cohort studies and thus lead already to a large number of insights documented in hundreds of (medical) publications. A considerable portion of these publications employ results from imaging data. However, the full potential of analyzing organ shapes, textures and spatial relations quantitatively is not exploited so far.

The Rotterdam study. A prominent example is the Rotterdam Study111, accessed: 1/31/2014, initiated in 1990 in the city of Rotterdam, in the Netherlands. Similar to later studies, it was motivated by the demographic change with more and more elderly people suffering from different diseases and their interactions. After the initial study involving almost 8,000 men and women, follow-ups at four points in time were performed—the most recent examinations took place in the 2009-2011 period. In the later examination cycles, also new individuals were involved leading to datasets from almost 15,000 patients Hofman2009 .

The original focus of the Rotterdam Study was on neurological diseases, but meanwhile it has been extended to other common diseases including cardiovascular and metabolic diseases. The study has an enormous impact on epidemiological and related medical research, documented in 797 journal publications registered in the pubmed database (search with keyword “Rotterdam Study”, January 30, 2014). Among them are predictions for the future prevalence of heart diseases and many studies on potential risk factors for neurodegenerative diseases. For a comprehensive overview of the findings, see Hofman2009 that summarizes the findings of more than 240 papers related to the Rotterdam Study. In a similar way, Hofman2011 is a significant update of these findings with more recent data.

Norwegian Aging Study. A long-term study in Norway investigates the relations between brain anatomy (as well as brain function), cognitive function, and genetics in normally aging people.222, accessed: 1/31/2014 In total 170 individuals (120 of them female), aged between 46 and 77 (mean 62), were examined in Bergen and Oslo in by now three waves (1st wave in 2004/2005, next in 2008/2009, and most recently in 2011/2012) Ystad2009 . While naturally not all of these subjects could be followed through all three waves, still most of them were subjected to an extensive combination of

  1. neuropsychological tests, including tests of the intellectual, language (memory), sensory/motor, and attention/executive function,

  2. MRI data, including co-registered T1-weighted anatomical imaging, diffusion tensor imaging, and – from the 2nd wave on – also resting-state functional MRI, as well as

  3. genotyping (1st wave only) YstadPhD .

The substantially heterogeneous imaging and test data are used to study aging-related questions about the modern Norwegian population, for example, how anatomical and functional changes in the human brain possibly relate to the later development of Alzheimer and dementia. Important findings include the relation between hippocampal volumes and memory function in elderly women Ystad2009 and the relation between subcortical functional connectivity and verbal episodic memory function in healthy elderly Ystad2010 .

SHIP. The Study of Health in Pommerania (SHIP) is another cohort study broadly investigating findings and their potential prognostic value for a wide range of diseases. The SHIP tries to explain health-related differences after the German reunion between East and West Germany. It was initiated in the extreme northeast of Germany, a region with high unemployment and a relatively low life expectancy.

In the first examination cycle (1997-2001) 4,308 adults of all age groups were examined, followed by a second and a third cycle that was finished at the end of 2012. The instruments used changed over time with some initial image data (liver and gallbladder ultrasound) available already in the first cycle and others, in particular whole body MRI, added later. The use of whole body MRI was unique in 2008 when the third examination cycle started. Breast MRI for women is performed, whereas for men MR angiography data are acquired, since men suffer from cardiovascular diseases significantly earlier than women Volzke2011 . In addition, a second cohort (SHIP-Trend) was established comprising 4,420 adult participants.

Diagnostic reports are created by two independent radiologists who follow strict guidelines to report their findings in a standardized manner. The pilot study to discuss the viability and potential of such a comprehensive MR exam is described by Hegenscheid2009 . The overall time for the investigation is two (complete) days with 90 minutes for the MR exam. The SHIP helped to reliably determine the prevalence of risk factors, such as obesity, and diseases. Major findings of the SHIP are increased levels of obesity and high blood pressure (compared to the German population) in the cohort. The MR exams alone identified pathological findings in 35% of the sample population. More than 400 publications in peer-reviewed journals are based on SHIP data (January 2014).

UK Biobank. The UK Biobank started recently and represents a comprehensive approach to study diseases with a high prevalence in an aging society, such as hearing loss, diabetes and lung diseases. Half a million individuals will be investigated in one examination cycle from which 100,000 receive an MRI from 2014 onwards. The rationale for the number of individuals to be included is explained by Peterson and colleagues Peterson2013 : they aim at a reliable identification of even moderate risk factors (RR between 1,3 and 1,5) for diseases with a prevalence of 5%. The prospective study should have a comprehensive protocol of cardiac MRI, brain MRI and abdominal MRI. This prospective cohort study also involves genetic information.333, accessed: 1/31/2014

The German National Cohort. The recently started “German National Cohort” in Germany is based on experiences with a number of moderate-size studies, such as SHIP, and examines some 200,000 individuals over a period of 10-20 years. Individuals will be invited in three waves to characterize changes. Due to the large-scale character, imaging is distributed over five cities. Thus, the subtle differences in imaging within different scanners have to be considered. 444, accessed: 1/31/2014 It explicitly aims at improvements in the treatment of chronic diseases and involves a variety of tissue samples, e.g., lymphocytes. Imaging in 30,000 individuals is again performed with MRI, comprising whole body, brain and heart.

2.4 Epidemiological Data

Epidemiological data are huge and very heterogeneous. As an example, in the UK biobank 329 attributes relate to physical measures, such as pulse rate, systolic and diastolic blood pressure, and various measures relate to vision or hearing. 471 attributes relate to interviews (socio-demographics, health history, lifestyle, …).

The data that are stored per individual is standardized but not completely the same, e.g., childbirth status and menstrual period are available for women only. Image data and derived information, e.g., segmentation results, significantly increased both the amount and complexity of data. Longitudinal cohort study data are time-dependent. While some instruments, such as blood pressure measurements, are available for all examination cycles, others were added later or removed. Individuals drop out, because they move, die or just do not accept the invitation to a second or third examination cycle. It is important to consider also such incomplete data but to be aware of potentially misleading conclusions.

The great potential of image-centric studies is that image data and associated laboratory data as well as data from interviews are available. An epidemiological study, such as the SHIP, has a large data dictionary

that precisely defines all attributes and their ranges. While laboratory data are scalar values, most data from interviews are nominal or ordinal values. In particular, data from interviews exhibit an essential amount of uncertainty. Self-reports with respect to alcohol and drug use, cigarette smoking and sexual practices may be biased towards “expected or socially accepted” answers. Epidemiologists are not only aware of these problems but developed strategies to minimize the negative effects, e.g., by asking redundant questions. After data collection, experts spend a lot of effort to improve the quality of the data. Despite these efforts, visual analytics techniques have to consider outliers, missing and erroneous data.

Geographic data. Geographic data play a central role in public health where the dynamics of local infections are visualized and analyzed (disease mapping). Chui and colleagues Chui:2011gm presented a visual analytics solution directly addressing this problem by combining three dedicated views. Also in cohort study data, geographic data are potentially interesting to understand local differences in the frequency and severity of diseases as an interaction between environmental factors and genetic differences. This branch of epidemiology is referred to as spatial epidemiology. Beale and colleagues Beale:2008dz investigated differences between rural and urban populations. In their comprehensive survey, Jerrett and colleagues Jerrett:2010 considered spatial epidemiology as an emerging area. However, we do not focus on spatial epidemiology since cohort study data typically comprise rather narrow regions and thus may not fully support such analysis questions.

2.5 Analysis of Epidemiological Workflow

The following discussion of observations and requirements for computer support is largely based on discussions with epidemiologists as well as the inspiring publication by Thew and colleagues. According to Thew:2009vz

  • epidemiological hypotheses are mostly observations made by physicians in clinical routine,

  • corresponding attributes are chosen based on the observations and further experience, and

  • regression analysis is frequently used to determine whether the investigated attribute is a risk factor or not.

Major requirements for an epidemiological workflow (again based on Thew:2009vz ) are:

  • Results have to be reproducible. Due to the iterative data assessment, methods need to be applied to new data sets as well and the results need to be comparable between different assessment times to characterize the change. User input needs to be monitored all the time to enable reproducible results.

  • A major result of an epidemiological analysis is whether certain factors influence a disease significantly. Relative risk (as a measure of effect size) and p-values as statistical significance level are particularly important.

Although these requirements neither consider image data nor visual analytics, they have to be considered also in these more innovative settings. Reproducibility, for example, means that clustering with random initialization is not feasible. Moreover, reports must be generated that clearly reveal all settings, e.g., parameters of clustering algorithms that were used for generating the results.

Since statistical analysis plays such an important role, statistics packages, such as SPSS555, accessed: 1/31/2014, R666, accessed: 1/31/2014 and STATA777, accessed: 1/31/2014

dominate in epidemiology. They provide various statistical tests also in cases where assumptions, such as a normal distribution, are not valid. Also the peculiarities of categorical data are considered. Visual analysis, so far, plays a minor role. As an example, Figures 

1 and 2 illustrate two graphical representations frequently used in epidemiology: Kaplan-Meier curves and interaction terms. A Kaplan-Meier curve shows the survival of patients, often as a comparison between different treatment options.


Figure 1: A Kaplan-Meier curve indicates how many patients survive at least a certain time. The more patients pass away, the larger is the confidence interval indicated by the dotted lines. The crosses mark each time a patient dies to further provide information on the reliability of the data that decreases over time (Courtesy of Petra Specht, University of Magdeburg).


Figure 2: The relative risk for cholelithiasis in men associated with a high level of a certain type of cholesterol slightly increases with a low BMI, but decreases for individuals with high or very high BMI. This multifactorial situation is depicted in an interaction term (Inspired by Volzke2005 ).

3 Visual Analytics in Epidemiology

The visualization of correlations in the epidemiological routine is largely restricted to scatterplots with regression lines and box plots to convey a distribution. Scatterplots may be enhanced, e.g., by coloring items according to certain characteristics, e.g., a diagnosis or by adding results from cluster or Principal Component Analysis (PCA)

Steenwijk:2010wo . The frequency related to a particular combination of values is often encoded by adapting saturation or darkness of colors.

Visual analytics methods can complement the statistical analysis and provide methods to explore the data. Efficient methods are essential to cope with the large amount of data and provide rapid feedback that is essential for any exploration process.

One of the first attempts to employ information visualization techniques for medical (image and non-image) data was realized in the WEAVE system Gresh:2000 . The system incorporated parallel coordinates as well as real time synchronization between different views. Another essential tool, inspired by the WEAVE system, was presented by Blaas and colleagues Blaas2007 . They enabled feature derivation techniques and incorporated segmentation techniques providing a powerful framework for heterogeneous medical data. Later work by Steenwijk and colleagues was more focussed on epidemiology. They provided an exploratory approach to analyze heterogeneous epidemiological data sets, including MRI Steenwijk:2010wo . They consider parameters on normalized and not normalized domains, while only normalized domains are comparable between subjects. Normalization means, for example, to register MRI brain data to an atlas to compare individual differences.


are used to project data into normalized domains. As an example, in brain analysis, a mapper defines the relation between an individual brain and a brain atlas that contains normalized and averaged information derived from many individual data. Feature extraction pipelines can be build visually by using a pipeline of mappers. The visualization is realized through multiple coordinated views which either represent scalar data or volumetric images. Different techniques to color code data, to align them and add further information are provided to enhance scatterplots. Steenwijk and colleagues evaluated their tools with specific examples from neuroimaging and questions related to a neurological disease where relations between clinical data (anxiety-depresssion scales, mental state scales) and MR-related data are analyzed (Fig. 

3). Normalized data domains are represented using scatterplots and parallel coordinated views. Dynamic changes are visualized using a time plot. The selection is linked between views and allows for multi-parameter comparison of clusters.

Zhang and colleagues Zhang:2012uw build a web-based information visualization framework for epidemiological analysis through different views. They divide the analytics process in batch analytics and on-demand analytics. Batch analytics steps are performed automatically as a new subject is added to the data set and aim to create groups by means of a certain condition. On-demand analytics are performed by user requests. Subjects are visualized using treemaps, histograms, radial visualizations and list views. However, neither filtering and grouping nor the interaction between the views are explained.

Recently, Turkay and colleagues Turkay:2013 described a framework to analyze the data of the Norwegian aging study, aiming particularly at hypothesis generation

. For this purpose, they give an overview on the dimensions in their dataset that conveys statistical properties, such as mean, standard deviation, skewness and kurtosis. The two latter measures characterize how asymmetric the data distribution is. Scatterplots display pairwise measures related to

all dimensions. Deviation plots, a new technique, enables grouping and supports a comparison of measures for a subgroup to the whole cohort.

The possibilities of VA tools can be summarized as follows:

  • Manual/automatic definition (brushing) of interesting parameters and ranges of values in attribute views,

  • Linking of attribute views for identifying relations,

  • Analysis across aggregation levels, parameters and subjects,

  • Definition of groups either interactively by means of (complex) brushes, or semi-automatically by means of clustering, and

  • Visual queries and direct feedback enable easy exploration

Figure 3: Left: A scatterplot relates magnetic transfer ratios to age. Items relating to the same patient (over time) are connected via a line. Colors indicate a diagnosis. Right: An enhanced scatterplot with PCA results for three subgroups overlaid (Courtesy of Martijn Steenwijk, VU University Medical Center Amsterdam).

While having different applications in mind, the Polaris system of Stolte and colleagues also employs multiple coordinated views to validate hypotheses Stolte:2002s . It uses a variety of different information visualization techniques to map ordinal/nominal or quantitative data of a relational data base. The system itself formulates data base queries and the mapping to create visualizations for the requested attributes. They choose the visualization mapping automatically based on the attribute types that are viewed in context with each other. This allows for a fast visualization of different attribute combinations in order to drill down to the information of interest. General visual analytics tools, such as Polaris and Weaver, in principle support some of the requirements for epidemiology. The use of coordinated views, brushes and switches is advantageous Weaver:2010ur . However, they are not designed to cope with the special requirements of cohort study data and do not directly support epidemiological workflows. In particular, no support for a combined assessment of image data and other epidemiological data is available.

Commercial Visual Analytics Systems.

There are a number of commercial software tools specialized on data visualization. As an example, Tableau

Mackinlay:2007jp provides an interface for creating different visualizations based on attribute drag-and-drop.888, accessed: 1/31/2014 These approaches deliver fast results with respect to visualizing different attributes. However, it is not supported to derive new data, such as scores for diseases, body mass index and other data that is relevant in epidemiology. QLikView is a similar tool for visualizing data associations. The user can design a frontend where associations can be assessed using multiple information visualizations. Thus, the user can drill down to the desired information. Statistics features regarding epidemiological key figures are limited.

Spotfire/IVEE Ahlberg:1996ep is able to handle more complex analysis of data sets and allows for interactive filtering of attributes. It can be linked to the statistical computing programming language R, which makes it versatile in comparison with its competitors. However, users need to be familiar with the R syntax.

Commercial systems cannot be enhanced or embedded in another system with hassle-free data exchange. The focus of commercial data visualization tools is business intelligence yielding a focus on quantitative data sources. At the same time they excel at incorporating user collaboration by including comment sections and share filters or entire setups of a dashboard.

4 Analysis of Medical Image Data for Epidemiology

Medical images are not by themselves useful for epidemiological analysis, since the semantics of image elements (pixel, voxels) is too low. The resulting extremely high-dimensional feature space would be unsuitable for visual analysis. Hence, image data is sequentially aggregated and reduced. The different steps of this process i.e. image analysis, shape analysis of extracted objects and subsequent clustering are characterized and discussed in this section. Throughout the section, we often refer to one case study, where MRI data from the lumbar spine is analyzed. The representation of assumptions on the lumbar spine shape and location as well as the object detection scheme used are examples of viable and common approaches. We do not claim that these techniques are better than any other approaches.

4.1 Medical Image Analysis

One of the major purposes of image analysis for a cohort study is to quantify anatomy, e.g., by volume, shape or spatial relations between structures. Quantification may be used to establish a range of normal values for different age groups and to characterize variations. Such variations may also confirm a disease and thus add to data derived from clinical tests. Thus, the use of image data enables more reliable conclusions w.r.t. the incidence and prevalence of diseases. As discussed in Sect. 2.2, MRI is particularly interesting because of the wide range of different information represented in these data. Thus, the examples discussed in the following relate to MRI data although the principles are more general.

Due to the wide range of image analysis tasks, the techniques should be adaptable to different analysis goals. The parameterization of an image analysis technique should be intuitive and the interaction should be kept to a minimum. The latter aspect is particularly important, since often several thousand datasets need to be analyzed. While largely manual approaches are acceptable in some clinical settings, such as radiation treatment planning, they are not feasible in the evaluation of cohort study data. The reduction of interactivity is not only a matter of effort, but also to meet the essential goal of reproducible approaches.

Detection and segmentation of anatomical structures. A modular system is a possible means to meet the central requirement of image analysis in cohort study data. An example of a cohort study is the liver segmentation of Gloger2010 , where concurrent detection and localization processes are combined for initial segmentation that is then fine-tuned in a model-driven segmentation step and finalized by a data-driven correction process. It has been shown that processes can be re-used and re-combined to solve a different segmentation task on similar data (kidney segmentation in MRI Gloger2012 ). Alternatively, the necessary domain knowledge, related to expected size, basic shape, position and grey values, can be separated from the detection and segmentation module. This strategy is attractive, since the user has not to care about the detection process when changing the application. Two problems have to be addressed in this case:

  • What is the expectation about the data support integrated to fit a model?

  • How is the with-class variation of the object in search separated from the between-class variation?

Point distribution models (PDM) McInerney1996 address the second question by training on sample segmentations. Model fitting is realized by a registration step. When training is not feasible, a prototypical model may be used instead. It is associated with restricted input about variation (a few parameters only) and qualitative knowledge about configuration or part-relationship.

In the following, we describe the linear elastic deformation of a finite element model (FEM) as a common method to model shape variation. The user specifies the average shape and two elasticity parameters: Young’s modulus defines how much external force is needed for a deformation and Poisson’s ratio describes how the deformation is transferred orthogonal to the direction of an incident force Rak2013 . The decomposition of the prototypical shape into finite elements bounded by nodes and specification of the elasticity parameters results in a stiffness matrix that relates the node displacement to incident forces (Eq. 1):


Different kinds of nodes may be specified that are attracted by different kinds of forces. Boundary nodes are attracted by the intensity gradient and inner nodes are attracted from expected intensity or texture. For letting an FEM move and deform into an object in an image, deformation is made dependent on time . Behavior then also depends on mass of the FEM and object-specific damping (Eq. 2).


represents the resistance of the moving FEM to external forces and allows the model to move over spurious image detail (e.g., gradients caused by noise). Damping avoids oscillation of the FEM. The system of differential equations is decoupled by solving the following generalized eigenproblem.


is the diagonal matrix with real-valued eigenvalues and

is the identity matrix.

Figure 4: Vibration modes 7 to 9 of a lower spine model. Vibration modes 1 to 6 represent rigid transformations (Courtesy of Marko Rak, University of Magdeburg).

After projecting the data on the eigenvector matrix, the differential equations can be solved fast and in a stable manner. Moreover, the projection on the eigenvectors (called

modes of vibration, see Fig. 4) separates deformation into components representing rigid transformation, major deformation modes and remaining minor deformation modes. The vibration modes can be used similar to the variation modes of an ASM to derive a quality-of-fit formulation for a fitted model instance. Since only a few anatomical objects have such a specific shape that it can be described by a simple deformation model and since training of additional information should be avoided, it is useful to complement simple deformation with pre-specified information on part-relationships. Part-relationships may describe the configuration of the object of interest w.r.t neighboring structures or may represent decomposition into parts (see Petyt1998 ).

Extending the FEM to a hierarchical model requires the introduction of a second layer FEM. Each sub-shape is represented by an FEM on the first layer and sub-shape FEMs are connected to the second layer. The type of connection regulates dependencies between sub-shapes and may range from distance constraints to co-deformation. FEMs for the first and second layer are created and assembled in the same fashion than elements are assembled for the sub-shape FEM Rak2013 .

Case study: Analysis of vertebrae. Back pain and related diseases exhibit a high prevalence and are thus a focus of the SHIP (recall Volzke2011 ). Specific goals are

  • to define the prevalence of degenerative changes of the spine,

  • to identify risk factors for these changes,

  • to correlate degenerative changes with actual symptoms, and

  • to better understand the progress from minor disease to a severe problem that requires medical treatment.

Epidemiologists hypothesize that smoking, heavy physical activity and a number of drugs that are frequently used are risk factors for back pain. Based on clinical observations, epidemiologists suggested to focus on the lumbar spine – the lowest part of the spine comprising five vertebrae. As a first step, the spine and lumbar vertebrae should be detected in T1-weighted and T2-weighted MRI data from SHIP.

Although local optimization could be complemented by stochastic global optimization Engel2010 , Rak and colleagues used only local optimization, since the initialization is simple for the given data. The user places a model instance in a sagittal view on the middle slice of the image sequence which is then transformed based on local image attributes. The model is constructed according to the appearance of vertebrae and spine in a sample image sequence. Vertebra sub-shapes were connected with a spine sub-shape by a structural model on the second level.

The spine model supported proper localization of the vertebrae. Since its most discriminate aspect was the cylindrical shape, it was represented by a deformable cylinder consisting of inner nodes only. The vertebrae shape was represented indirectly by inner nodes as well, since reliability of the intensity gradient was low. For each of the two shape models, the vertebra and the spine, a weighted combination of the T1-weighted and the T2-weighted image was computed as appearance input. Weights for each of the two models were determined a priori and produced a clearly recognizable local minimum for vertebra and spine appearance, respectively. The user placed a model instance in the vicinity of the object on a sagittal slice. Computation time until convergence was between 1.1 and 2.6 seconds per case. The method was evaluated on 49 data sets from the SHIP. The detection was considered successful if the center of each vertebra sub-shape was in the corresponding vertebra in the image data, which was achieved in 48 of the 49 cases (see Fig. 5 for examples and Rak2013 for further details).

Figure 5: Examples for initialization and convergence of model instances applied to the MRI data (Courtesy of Marko Rak, University of Magdeburg).

4.2 Shape Analysis

Epidemiologists are used to work with numerical and categorical data which then is tested for statistic validity. Medical image data also allows to consider characteristic object shapes. As an example, the shape of the liver may depend in a characteristic manner on infections (hepatitis), alcohol consumption, or obesity. Eventually, shape characteristics may change even before a disease becomes symptomatic. If this turns out, shape changes may be employed as an early stage indicator.

While the quantitative analysis of shapes or parts thereof (morphometrics) is a recent trend in epidemiology due to the availability of image data, it is established in anatomy and evolutionary biology.

Shape analysis requires that different shapes are transformed in a common space, typically by a rigid transformation (translation and rotation). The parameters of this rigid transformation are determined in an optimization process that minimizes the distances of corresponding points. A major challenge is to determine these corresponding points that serve as landmarks. In particular for soft tissue structures there are not sufficient recognizable landmarks and therefore a parameterization is necessary to define these points. Without going into detail, we assume that this process is applied to many individual shapes , say livers in a cohort study. Then, for each an optimal non-rigid transformation to a reference shape defines a deformation with displacement vectors for each landmark.

For use in epidemiology, a large set of displacement vectors is not the right level of granularity. Instead, a few dimensions are desirable that characterize major differences. Thus, typically a dimension reduction technique, such as PCA, is employed to characterize the directions that represent the major differences. This process may be adapted to specific analysis questions by assigning individual weights to the landmarks expressing a strong interest in particular displacements Hermann2011 . Thus, epidemiological hypotheses may be incorporated.

While the establishment of point correspondences is often a major challenge, recently alternative approaches were developed. The GAMES algorithm (Growing and adaptive meshes) Ferrarini2007

creates a data structure to represent the shape variance if no pairwise correspondence between points is given. However, it can be prone to errors since it requires a prior registration of segmentation masks.

Shape analysis, of course, may also be supported by appropriate visualizations that enable pairwise comparisons and emphasize differences. In this vein, Busking2010 presented a system for shape space exploration based on carefully designed multiple coordinated views.

4.3 Analysis of Lumbar Spine Canal Variability

In Sect. 4.1, we introduced a case study related to the analysis of the lumbar spine in cohort data and explained how the spine and the vertebrae are detected in MRI data from the SHIP. Here, we extend this discussion by the analysis of the spinal canal and non-imaging attributes related to back pain. In the SHIP, attributes related to back pain history, e.g., working habits, physical activities, size and weight, are available to identify and analyze potential correlations with findings from the MRI data. After careful discussions, we selected 77 attributes (60 are ordinal or nominal and 17 scalar) to investigate back pain Klemm_2014_BVM . Ordinal data are primarily results of multiple choice questions. The epidemiologists suggested to focus on the overall shape and curvature of the spine in that region instead of individual vertebrae. This overall shape is well characterized by the lumbar canal. Thus, correlations between the shape of the lumbar spine, attributes of back pain history and activities both in leisure and working time may be analyzed.


Figure 6: Lumbar spine visualization of 243 female subjects. Left: Tetrahedron-based finite element model from Rak and colleagues Rak2013 . The dashed purple line indicates the lumbar spine canal centerline Middle: Model used to detect lumbar spine canal in an MRI scan Right:

Agglomerative hierarchical clustering of 243 centerlines yields seven clusters. Their representatives are visualized as ribbons mapping cluster size to width. The ribbon color encodes the distance to the semi-transparent plane orthogonal to the view direction (lower inset). Shadow projections (upper inset) provide an additional visual hint on the curvature extent

Klemm2013 .

Klemm and colleagues Klemm2013 extracted, clustered and visualized spine canal centerlines. Image segmentation of 493 MRI data sets was carried out automatically using tetrahedron-based finite element models of vertebrae and spinal canal Rak2013 . Using barycentric coordinates of the tetrahedrons, a centerline consisting of 93 discrete points was extracted for each segmentation, as seen in Figure 6 (left, middle). They served as input for an agglomerative hierarchical clustering which created groups of subjects based on differences in shape. This special clustering technique was chosen, since it produces meaningful results in the clustering of similar structures, such as fiber tracts derived from MR-Diffusion Tensor Imaging data Klemm2013 .

The cluster visualization in Figure 6 (right) displays each cluster representative as ribbon in a sagittal plane. The representative is the centerline with the smallest sum of distances to all other centerlines, i.e., the centroid line of the cluster. The width of the ribbon encodes the cluster count and the color encodes the distance to the sagittal plane. Shadow projections also (redundantly) convey the distance to the sagittal plane. This allows to assess the 3D shape in a 2D projection. The results of the clustering can be used in different ways.

  • Outlier detection: Extraordinary shapes yield clusters of small size that differ strongly from the global mean shape. This can point to pathologies or errors in the segmentation process.

  • Hypothesis generation: Shape groups serve as a starting point for an exploratory analysis to analyze disease-related correlations. The usual workflow requires epidemiologists to define groups, e.g., age ranges. Groups calculated solely using shape information can be analyzed to detect statistically relevant associations in other expositions which can lead to new hypotheses.

  • Hypothesis validation: Clustering based on non-image related features can be used to analyze if these clusters are correlated to characteristic shapes, e.g., a strongly bent spine canal representative.

Calculating curvature on groups created according to body height starting from 150 cm in 10 cm steps was performed. Klemm and colleagues found that taller people have a more straight spine compared to small people. They also found multiple clusters of people 10 years above average age across all groups that exhibit a strong ”S” shape of the spine, which was the starting point for new investigations using expert chosen spine-related attributes. This method was extended to integrate the relevant information for identifying correlations. Thus, for a selected cluster, information related to the distribution of attributes, such as the back pain history (frequency and intensity of back pain), may be displayed as a tool tip. The initial observations show, that a box plot summarizing the distribution is more suitable than the full histogram. For routine use in epidemiology, the lumbar spine visualization (Fig. 6) has to be complemented with at least simple statistics to answer questions, such as: Is there a statistically significant difference between the curvature of the lumbar spine canal and back pain frequency? If so, what is the effect size?

4.4 Cluster Analysis and Information Space Reduction

A crucial task in epidemiology is the definition of groups of subjects. Differences and similarities among groups are investigated and control groups are defined to detect and assess the impact and interaction of risk factors to define the relative risk. A straightforward approach is the manual definition based on study variables and ranges of interest. A data-driven extension is the automatic detection of potentially relevant subgroups in the often high-dimensional data by means of clustering algorithms Klemm_2014_BVM . In particular, the generation of new hypotheses, which may be tightly connected to the identification of new groups, benefits from the latter. In clustering, subjects, being similar with respect to a certain similarity metric, are grouped in clusters with a low intra-cluster and a high inter-cluster variance. Particular challenges in the cluster analysis of cohort study data are:

  • Missing data, e.g., denied answers to inconvenient questions Donders2006

  • Mixture of scala and categorical study variables AhmadD07

  • Time-varying variables in longitudinal studies Genolini2010

As a consequence of the first problem, it should be reported to the user how many datasets were actually used for clustering. Depending on the chosen attributes, this may be a subset of the overall amount of data. Incomplete datasets may and should be used when the relevant data is available. It is essential to use similarity metrics that consider also ordinal or categorical data. Usually, the following convention is used: The distance between datasets equals 1 if their categorical data is different and 0 if it is the same. With ordinal data, more care is necessary. The difference between ”strongly agree” and ”strongly disagree” on a Likert scale is larger than the difference between ”agree” and ”disagree”. However, the precise quantification is not straightforward. As a first step to explore the SHIP data, a parallel coordinate view is combined with scatterplots and clustering (Fig. 7).

Figure 7: A parallel coordinate view enables the selection of a relevant subset (here persons with a weight larger than 120 kg). The scatterplots represent correlations between age, size and weight. The elements are color-coded according to a clustering result that yields three clusters. The encircled elements correspond to the selection in the parallel coordinates view.

With respect to clustering algorithms, it is essential that the number of resulting groups has not to be specified in advance. Moreover, algorithms are preferred that allow outliers instead of forcing all elements to be part of a cluster. Outliers may be particularly interesting and thus serve as a starting point for further investigation. Of course, they may also indicate a bad quality of some data. In our experiments, density-based clustering with DBSCAN Ester:1996 produced plausible results when applied to non-image data of the SHIP study. The DBSCAN result is sensitive to the minPoints parameter that determines the minimum size of a cluster. Some cycles of clustering are necessary to make a suitable choice. In this process, an appropriate visualization is essential to easily understand the results. A visualization that conveys the location, size and shape of clusters is difficult in case that clustering is applied to more than two-dimensional data. A recently developed approach enables 3D visualizations of clustering results with very low levels of occlusion Glasser_2014_GRAPP .

The majority of cohort-studies are restricted to non-image data, i.e., categorical and scalar data, which may directly serve as input for a clustering algorithm. In Axen2011 , 176 patients with lower back pain have been monitored over six months via text messages describing their bothersomeness. All patients received chiropractic treatment. A hierarchical clustering of the individual temporal courses of bothersomeness revealed groups of patients who responded differently to the therapy, which may improve the optimal individual treatment selection. Hypotheses about the differences between paternal age-related schizophrenia (PARS) and other cases of schizophrenia were generated in Lee2011

. A k-means clustering of demographic variables, symptoms, cognitive tests and olfaction for 136 subjects (34 with PARS) delivered clusters containing a high concentration of PARS cases. Significant characteristics of these clusters may give a hint on features of PARS improving its dissociation of other cases of schizophrenia.

In analyzing image-centric cohort study data, besides categorical and scalar data, images may be clustered for group definition. Image intensities, segmentation results, e.g., the surface of the segmented liver or the centerline of the spinal canal (recall Section 4.3) and derived information, e.g., liver tissue texture, liver volume and spinal canal centerline geometry, may serve as input. In Seiler2012 , the anatomical variation of the mandibles is assessed across a population. For treating mandible fractures, subjects with similar characteristics are grouped in clusters and a suitable implant is designed per cluster. The clustering algorithm k-means is applied to transformation parameters of a locally affine registration between all mandible surfaces segmented in CT data. A cohort of 50 patients with suspicious breast lesions was investigated by means of dynamic contrast enhanced MRI (DCE-MRI) in Preim2012 . Each lesion was clustered according to its perfusion characteristics by means of a region merging approach. Perfusion is represented by the temporal course of the DCE-MRI signal intensities. The clustering itself did not generate groups of patients here, but based on each individual number of clusters and their perfusion characteristics, two groups of lesions could be defined: benign and malignant. These groups were then compared to histological results from core needle biopsy.

Investigating high-dimensional non-image cohort study data often benefits from an information space reduction, e.g., by means of PCA. Plotting the data for inspection in a lower dimensional space while capturing the greatest level of variation, e.g., a scatterplot of the first two principal components, as well as detecting trends in the data and ordering them according to the variance they describe are important applications of PCA. In Robertson2008 , symptom data gathered in interviews of 410 people with Turret syndrome was investigated in order to specify homogeneous symptom categories for a better characterization of the disease’s phenotype. First, clusters of symptom variables were generated using agglomerative hierarchical clustering. Then, for each cluster and each participant a score was computed equal to the sum of present symptom variables in the cluster. These scores were the input for PCA, which produced homogeneous symptom categories, sorted according to their percentage of represented symptomatic variance. In Steenwijk:2010wo , scatterplots of variables from cohort study data are extended by superimposing PCA ellipses. The ellipses are computed per group of subjects and illustrate its global distribution with respect to the two opposed variables. They are spanned by the principal component axes of a groups data points and centered at their mean. Optionally, their transparency is adjusted with respect to the groups confidence, i.e. the number of contained subjects.

4.5 Categorical Data

Cohort study data sets comprise many categorical data such as answers to questionnaires or categorizations that may result from binning continuous data, such as intervals of income or age groups. These data are discrete and exhibit a low range. While data resulting from binning or from answers marked at a Likert scale have an inherent order (ordinal data), often no inherent order exists. Standard information visualization methods like scatterplots and parallel coordinates are designed for continuous data and thus not ideal for displaying categorical data, since many data points occlude each other.

Figure 8: Parallel sets are designed to explore categorical data. Parallel sets are based on the parallel coordinates layout, but it maps frequency of the data instead of rendering them just as data points. The user can interactively remap the data to new categorizations as well as highlight entries to examine their distribution along other mapped dimensions. The displayed data set shows the distribution of passengers of the RMS Titanic and whether they survived the sinking of the ship Davies2013 .

Categorical data are often visualized using boxes where the width is scaled to frequency Wittenburg:2001 . These approaches use much space and are also not well suited for encoding multidimensional relationships. Parallel sets, introduced by Bendix and colleagues, comprise the same layout as parallel coordinates, ”but the continuous axes were replaced with sets of boxes …scaled to the frequency of the category” Bendix:2005vf (see Fig. 8). Selecting an attribute will map each category to a distinct color so that they can be traced through all visualized dimensions. Highlighting a category may be realized by drawing the selected category in a higher saturation leading to a pop-out effect. It is also useful to display a histogram for the selected category annotated with statistical information. Selecting a category will only display the particular box on an axis and make more room for connections to other axes. This is especially helpful if the number of categories for one dimension is large. In case of data without inherent order, reordering of axes is possible and supports the exploration by providing more comprehensible layouts.

5 Concluding Remarks

A variety of large prospective cohort studies are established and ongoing. They generate a wealth of potentially relevant information for assessing risks, for estimating costs involved in treatment and thus inform health policy makers with respect to potentially preventive measures or cost-limiting initiatives. Despite the great potential of data mining and interactive visualization, none of these studies included such activities in their original planning. The role of computer science, so far, was limited to database management and data security. Visual analytics has a great potential for exploring complex health-related data, as recently shown by Zhang and colleagues Zhang:2013 for clinical applications, such as treatment planning. Similar techniques may be employed to address epidemiology research. In contrast to clinical applications, where a severe time pressure leads to strongly guided workflows, epidemiology research benefits from powerful and flexible tools that enable and support exploration. Currently, existing and widespread information visualization and analytics techniques are employed and adapted to epidemiologic data. The high-dimensional nature of these data, however, also requires to develop new techniques.

The specific and new aspect discussed in this paper was the integration of image data, information derived from image data, such as spine curvature-related measures, and more traditional socio-demographics data.

Future work. So far, visual analytics research and software was rarely focused on epidemiology. Thus, to adapt visual analytics to epidemiology and to integrate the solutions with tools familiar to epidemiologists is necessary. In epidemiology, national studies are prevailing, which is due to the large amount of legislative conditions to be considered. International studies would enable to explore diseases with lower prevalence, subtypes of diseases that occur rarely, e.g., cancer in early age, and specific questions of spatial epidemiology. The SHIP Brazil study will provide such information. It was recently initiated to perform a study in Brazil according to the standards and experiences gained in the German SHIP study. A nasty but essential problem of image-based epidemiologic study is quality control. Research efforts are necessary to automatically check whether image data fully cover the target region, whether the alignment of slices, e.g., in heart imaging, is correct and whether severe artifacts appear, e.g., motion artifacts. This chapter discussed exciting developments related to the combined use of radiologic image and more classical epidemiological data. The next wave is already clearly recognizable: genetics information will be integrated in the search for early markers associated with risks for diseases.

We want to thank Lisa Fraunstein, David Kilias and David Perlich who supported our analysis of the SHIP data as student workers as well as Marko Rak who provided the detection algorithm for the vertabrae and Myra Spilopoulou for fruitful discussions on clustering and data mining (all University of Magdeburg). We thank Martijn Steenwijk for providing images from his work and Charl Botha for fruitful discussions. Matthias Günther (Fraunhofer MEVIS) explained us quality aspects of MR imaging in epidemiologic studies. This work was supported by the DFG Priority Program 1335: Scalable Visual Analytics. SHIP is part of the Community Medicine Research net of the University of Greifswald, Germany, which is funded by the Federal Ministry of Education and Research (grant no. 03ZIK012), the Ministry of Cultural Affairs as well as the Social Ministry of the Federal State of Mecklenburg-West Pomerania.


  • (1)

    Review: A Gentle Introduction to Imputation of Missing Values.

    Journal of Clinical Epidemiology 59(10), 1087 – 1091 (2006)
  • (2) Ahlberg, C.: Spotfire: an information exploration environment. SIGMOD Record 25(4) (1996)
  • (3) Ahmad, A., Dey, L.: A k-mean Clustering Algorithm for Mixed Numeric and Categorical Data. Data Knowl. Eng. 63(2), 503–527 (2007)
  • (4) Axén, I., Bodin, L., Bergström, G., Halasz, L., Lange, F., Lövgren, P.W., Rosenbaum, A., Leboeuf-Yde, C., Jensen, I.: Clustering Patients on the Basis of Their Individual Course of Low Back Pain Over a Six Month Period. B M C Musculoskeletal Disorders 12, 99–108 (2011)
  • (5) Beale, L.L., Abellan, J.J., Hodgson, S.S., Jarup, L.L.: Methodologic issues and approaches to spatial epidemiology. Environmental health perspectives 116(8), 1105–1110 (2008)
  • (6) Bendix, F., Kosara, R., Hauser, H.: Parallel sets: Visual analysis of categorical data pp. 133–140 (2005)
  • (7) Blaas, J., Botha, C.P., Post, F.H.: Interactive Visualization of Multi-Field Medical Data Using Linked Physical and Feature-Space Views. In: Proc. of EuroVis, pp. 123–130 (2007)
  • (8) Busking, S., Botha, C.P., Post, F.H.: Dynamic Multi-View Exploration of Shape Spaces. Computer Graphics Forum 29(3), 973–982 (2010)
  • (9) Chui, K.K., Wenger, J.B., Cohen, S.A., Naumova, E.N.: Visual analytics for epidemiologists: understanding the interactions between age, time, and disease with multi-panel graphs. PloS one 6(2), e14,683 (2011)
  • (10) Davies, J.: Parallel Set of the Titanic Data Set. (2012). [Online; accessed 30-January-2014]
  • (11) Engel, K., Toennies, K.D.: Hierarchical vibrations for part-based recognition of complex objects. Pattern Recognition 43(8), 2681–2691 (2010)
  • (12) Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In: Proc. of the Second International Conference on Knowledge Discovery and Data Mining (KDD), pp. 226–231 (1996)
  • (13) Ferrarini, L., Olofsson, H., Palm, W., Vanbuchem, M., Reiber, J., Admiraalbehloul, F.: GAMEs: Growing and adaptive meshes for fully automatic shape modeling and analysis. Medical Image Analysis 11(3), 302–314 (2007)
  • (14) Fletcher, R.H., Fletcher, S.W.: Clinical Epidemiology. Lippincott Williams & Wilkins (2011)
  • (15) Genolini, C., Falissard, B.: KmL: k-means for Longitudinal Data. Comput. Stat. 25(2), 317–328 (2010)
  • (16) Glas̈ser, S., Lawonn, K., Preim, B.: Visualization of 3D Cluster Results for Medical Tomographic Image Data. In: In Proc. of Conference on Computer Graphics Theory and Applications (VISIGRAPP/GRAPP), pp. 169–176 (2014)
  • (17) Gloger, O., J., K., Stanski, A., Völzke, H., Puls, R.: A fully automatic three-step liver segmentation method on LDA-based probability maps for multiple contrast MR images. Magnetic Resonance Imaging 28(6), 882–897 (2010)
  • (18) Gloger, O., Toennies, K.D., Liebscher, V., Kugelmann, B., Laqua, R., H., V.: Prior shape level set segmentation on multistep generated probability maps of MR datasets for fully automatic kidney parenchyma volumetry. IEEE Transactions on Medical Imaging 31(2), 312–325 (2012)
  • (19) Gresh, D.L., Rogowitz, B.E., Winslow, R.L., Scollan, D.F., Yung, C.K.: Weave: a system for visually linking 3-d and statistical visualizations, applied to cardiac simulation and measurement data. In: Proc. of IEEE Visualization, pp. 489–492 (2000)
  • (20) Hegenscheid, K., Kühn, J.P., Völzke, H., Biffar, R., Hosten, N., Puls, R.: Whole-body magnetic resonance imaging of healthy volunteers: pilot study results from the population-based SHIP study. Fortschritte auf dem Gebiet der R”øntgenstrahlen und der bildgebenden Verfahren (Röfo) 181 (8), 748–759 (2009)
  • (21) Hermann, M., Schunke, A.C., Klein, R.: Semantically steered visual analysis of highly detailed morphometric shape spaces. In: Proc. of IEEE Symposium on Biological Data Visualization (BioVis), pp. 151–158 (2011)
  • (22) Hofman, A., Breteler, M.M.B., van Duijn, C.M., Janssen, H.L.A., Krestin, G.P., Kuipers, E.J., Stricker, B.H.C., Tiemeier, H., Uitterlinden, A.G., Vingerling, J.R., Witteman, J.C.M.: The Rotterdam Study: 2010 objectives and design update. European Journal of Epidemiology 24, 553–572 (2009)
  • (23) Hofman, A., van Duijn, C.M., Franco, O.H., et al.: The Rotterdam Study: 2012 objectives and design update. European Journal of Epidemiology 26, 657–686 (2011)
  • (24) Jerrett, M., Gale, S., Kontgis, C.: Spatial Modeling in Environmental and Public Health Research. International Journal of Environmental Research in Public Health 7(16), 1302–1329 (2010)
  • (25) Klemm, P., Frauenstein, L., Perlich, D., Hegenscheid, K., Völzke, H., Preim, B.: Clustering Socio-demographic and Medical Attribute Data in Cohort Studies. In: Proc. of Bildverarbeitung fuer die Medizin (BVM) (2014)
  • (26) Klemm, P., Lawonn, K., Rak, M., Preim, B., Tönnies, K., Hegenscheid, K., Völzke, H., Oeltze, S.: Visualization and Analysis of Lumbar Spine Canal Variability in Cohort Study Data. In: Proc. of Vision, Modeling, Visualization (VMV), pp. 121–128 (2013)
  • (27) Lee, H., Malaspina, D., Ahn, H., Perrin, M., Opler, M.G., Kleinhaus, K., Harlap, S., Goetz, R., Antonius, D.: Paternal Age Related Schizophrenia (PARS): Latent Subgroups Detected by k-means Clustering Analysis. Schizophrenia Research 128(1–3), 143 – 149 (2011)
  • (28) Mackinlay, J., Hanrahan, P., Stolte, C.: Show me: automatic presentation for visual analysis. IEEE Transactions on Visualization and Computer Graphics 13(6), 1137–1144 (2007)
  • (29) Marathe, M., Vullikanti, A.K.S.: Computational epidemiology. Communications of the ACM 56(7), 88–96 (2013)
  • (30) McInerney, T., Terzopoulos, D.: Deformable models in medical image analysis: a survey. Medical Image Analysis 1(2), 91–108 (1996)
  • (31) Pearce, N., Merletti, F.: Complexity, simplicity, and epidemiology. International Journal of Epidemiology 35(3), 515–519 (2006)
  • (32) Petersen, S.E., Matthews, P.M., Bamberg, F., et al.: Imaging in population science: cardiovascular magnetic resonance in 100,000 participants of UK Biobank - rationale, challenges and approaches. J. Cardiovasc Magn Reson. 28, 15–46 (2013)
  • (33) Petyt, M.: Introduction to finite element vibration analysis. Cambridge University Press (1998)
  • (34) Preim, U., Glaßer, S., Preim, B., Fischbach, F., Ricke, J.: Computer-aided Diagnosis in Breast DCE-MRI-Quantification of the Heterogeneity of Breast Lesions. European Journal of Radiology 81(7), 1532–1538 (2012)
  • (35) Rak, M., Engel, K., Tönnies, K.D.: Closed-Form Hierarchical Finite Element Models for Part-Based Object Detection. In: Proc. of Vision, Modeling, Visualization (VMV), pp. 137–144 (2013)
  • (36) Robertson, M.M., Althoff, R.R., Hafez, A., Pauls, D.L.: Principal Components Analysis of a Large Cohort With Tourette Syndrome. The British Journal of Psychiatry 193(1), 31–36 (2008)
  • (37) Seiler, C., Pennec, X., Reyes, M.: Capturing the Multiscale Anatomical shape Variability with Polyaffine Transformation Trees. Medical Image Analysis 16(7), 1371 – 1384 (2012)
  • (38) Steenwijk, M.D., Milles, J., van Buchem, M.A., Reiber, J.H.C., Botha, C.P.: Integrated Visual Analysis for Heterogeneous Datasets in Cohort Studies. In: Proc. of IEEE VisWeek Workshop on Visual Analytics in Health Care (2010)
  • (39) Stolte, C., Tang, D., Hanrahan, P.: Polaris: A system for query, analysis, and visualization of multidimensional relational databases. IEEE Transactions on Visualization and Computer Graphics 8(1), 52–65 (2002)
  • (40) Thew, S., Sutcliffe, A., Procter, R., de Bruijn, O., McNaught, J., Venters, C.C., Buchan, I.: Requirements engineering for e-Science: experiences in epidemiology. IEEE Software 26(1), 80–87 (2009)
  • (41) Turkay, C., Lundervold, A., Lundervold, A.J., Hauser, H.: Hypothesis generation by interactive visual exploration of heterogeneous medical data. In: Proc. of Human-Computer Interaction and Knowledge Discovery in Complex, Unstructured, Big Data, pp. 1–12 (2013)
  • (42) Völzke, H., Alte, D., Schmidt, C., et al.: Cohort Profile: The Study of Health in Pomerania. International Journal of Epidemiology 40(2), 294–307 (2011)
  • (43) Völzke, H., Baumeister, S.E., Alte, D., Hoffmann, W., Schwahn, C., Simon, P., John, U., Lerch, M.M.: Independent Risk Factors for Gallstone Formation in a Region with High Cholelithiasis Prevalence. Digestion 71, 97–105 (2005)
  • (44) Weaver, C.: Cross-filtered views for multidimensional visual analysis. IEEE Transactions on Visualization and Computer Graphics 16(2), 192–204 (2010)
  • (45) Wittenburg, K., Lanning, T., Heinrichs, M., Stanton, M.: Parallel bargrams for consumer-based information exploration and choice. In: Proc. of the ACM symposium on User interface software and technology (UIST), pp. 51–60 (2001)
  • (46) Ystad, M.: Quantitative structural and functional brain imaging in cognitive aging. Ph.D. thesis, University of Bergen (2010)
  • (47) Ystad, M., Eichele, T., Lundervold, A.J., Lundervold, A.: Subcortical functional connectivity and verbal episodic memory in healthy elderly—resting state fmri study. NeuroImage 52(1), 379–388 (2010)
  • (48) Ystad, M., Lundervold, A.J., Wehling, E., Espeseth, T., Rootwelt, H., Westlye, L., Andersson, M., Adolfsdottir, S., Geitung, J., Fjell, A., Reinvang, I., Lundervold, A.: Hippocampal volumes are important predictors for memory function in elderly women. BMC Medical Imaging 9(1), 1–15 (2009)
  • (49) Zhang, Z., Gotz, D., Perer, A.: Interactive Visual Patient Cohort Analysis. In: Proc. of IEEE VisWeek Workshop on Visual Analytics in Healthcare (2012)
  • (50) Zhang, Z., Wang, B., Ahmed, F., Ramakrishnan, I., Viccellio, A., Zhao, R., Mueller, K.: The Five W s for Information Visualization with Application to Healthcare Informatics. IEEE Transactions on Visualization and Computer Graphics 19(11), 379–388 (2013)