sign in sign up
<<
Home , Radar chart, Statistical graphics

Paper PO10 Clinical Adverse Events Analysis and Shi-Tao Yeh, GlaxoSmithKline, King of Prussia, PA.

ABSTRACT

A collects the adverse events (AE) data during the trial as part of the clinical safety profile. There is a need for greater vigor in the analysis and presentation of the AE data. This paper discusses the approaches to construct the graphical models for displaying and analyzing AE data.

The graphical models described in this paper are:

1) bar for summary display, 2) dot plots for the most frequent AEs sorted by variables of interest, 3) a cluster , 4) a Kaplan-Meier (KM) time-to-event , 5) tree , 6) a , and 7) constellation diagrams.

The SAS v9 products used in this paper are SAS BASE®, SAS/STAT® SAS Enterprise Miner™ (EM) 5.2, and SAS/GRAPH® on the PC Windows® platform and on the UNIX environment.

KEY WORDS adverse event, , bar charts, dot plot, cluster chart, Kaplan-Meier plot, tree diagrams, radar chart, constellation diagrams.

INTRODUCTION In clinical trial studies, the investigators are responsible for recording a subject's adverse experiences (AEs) at each clinical office visit or assessment. An AE is defined as: any noxious, pathological or unintended change in anatomical, physiological or metabolic functions as indicated by physical signs, symptoms and/or laboratory changes occurring in any phase of the clinical study whether associated with the study drug, or active comparator or placebo, and whether or not such change is considered drug related.

The instructions for the investigator in the Case Report Form (CRF) are:

Have there been any adverse experiences observed or elicited by the following direct question to the patient: “Have you felt different in any way since the last assessment? If the answer is ‘Yes’, please record details in the Adverse Experiences section.

These AEs are stored in a clinical trial study . The reported AEs are part of clinical safety data. It is a challenging task for the study team to report and interpret the AE data. The traditional study AE reports tend to consist of basic summary tables and listings. It is difficult to detect AE patterns in a presentation. Substituting graphs for tables increases the of review. This paper applies statistical methods, such as predictive models, multivariate analysis, , and data mining techniques, for AE reporting. The graphical models described in this paper are:

1) bar charts for summary display, 2) dot plots for the most frequent AEs sorted by variables of interest, 3) a cluster chart, 4) a Kaplan-Meier (KM) time-to-event plot, 5) tree diagrams,

1 6) a radar chart, and 7) constellation diagrams. Some of the graphical presentations also provide interactive features. Interactive graphs allow for user (human) interactions. The users interact directly with the graphs, mainly through user interfaces. User interfaces are controlled with a mouse as well as on screen buttons and menus. The interactive graphs provide the following desirable features. The user can:

1) change graphical properties, including modification of titles and axes through the interactive menus, 2) use the data tips feature to include extra on the graph, 3) link to subsequent graphs, or tables, 4) change the scale using zooming techniques. Zoomable interfaces come in two types: a. Geometric zooming where the scale is linear with the multiplier. b. Semantic zooming where objects may change shape at different zoom levels. Semantic zooming is a distortion technique that displays the object in a fisheye view. By moving around and changing the scale, the user can customize the data display; and 5) produce slide show of static images

The SAS System implements several new technologies such as Java applets and ActiveX Control to produce the interactive graphs. SAS macros, DS2TREE and DS2CONST are selected for this application. SAS macros DS2CONST and DS2TREE use the Constellation Applet and the Treeview Applet respectively, to produce the node-link Constellation and Treeview diagram. Each node can be linked to one or more other nodes. Unlike the Treeview Applet, the Constellation Applet does not require a hierarchical relationship between the nodes. Although it can be used to display hierarchical relationships, the Constellation Applet does not automatically place the root node at the center of the display.

Some of the graphical presentations use SAS ODS Graphics. If there is a need to modify the ODS Graphics, it is difficult to accomplish this task. The appearance of the ODS Graphics is governed by the ODS Graphical Template Language (GTL). The modification of ODS Graphics can be made from three sources: 1) from the that is used for the procedure execution, 2) from changing the style template, and 3) from modification of the graph template. In most cases, it involves the modification of the ODS Template or the need to create a new ODS Graphics Template.

A simulated hypothetical clinical AEs data set has been used throughout the paper for purposes only.

The SAS v9 products used in this paper are SAS BASE®, SAS/STAT® SAS Enterprise Miner™ (EM) 5.2, and SAS/GRAPH® on the PC Windows® platform and on the UNIX environment.

I. BAR CHARTS FOR AE DISPLAY

A is a bivariate display summarizing a set of categorical data. It is often used in exploratory to illustrate the major features of the distribution of the data in a rectangular bar format. It displays the data using a number of rectangles, of the same width, each of which represents a particular category. The size of each bar is proportional to the number of cases in the category it represents. It can be displayed horizontally or vertically. A brief sample AE summary table is shown below for illustration purposes. A complete AE summary table may contain dozens of page.

Table 1. Sample AE Summary Table

2 The AE summary table shown in Table 1 can be presented in the bar charts display shown in Figure 1.

Figure 1. Bar Chart Display of AE by Intensity

We can add an additional category – the time trend and patterns of the first AE occurrence. Figure 2 shows the display of time to the first AE occurrence.

Figure 2. Sample AE Display of Time to First Occurrence

3 II. DOT CHARTS

A dot plot is a primitive display of a set of data with symbols, usually circular dots. For nominal or , a dot plot is similar to a bar chart, with the bars replaced by a series of dots. Each dot represents a summary statistics value. For continuous data, the dot plot is similar to a , with the rectangles replaced by dots. The dot plot display needs to have symbols arranged in their proper locations on a scale representing summary statistics values. The horizontal axis represents the scale and vertical axis is for categorical listing. It can be sorted by variables of interest. A dot plot can also help detect any unusual observations or depict distributions in detail. The summary statistics in Table 1 can be displayed as dot plot shown in Figure 3 below.

It displays the AE preferred term on the y-axis and one summary statistics, either summation of event count (N) or percentage (%), as a scale on the x-axis. Figure 3 also utilizes the left hand space to display the summary statistics of summation count N, and percentage PCT (%).

Figure 3. Sample AE Dot Plot

4 Figure 4 uses different symbols for different treatments and overlays symbols for easier comparison. It is also a design with multiple graphs on one page. The right hand panel displays a relative risk plot, graphically presenting the relative risks, with associated 95% conference intervals, these helping to determine risk relative to other signals.

Figure 4. Sample AE Dot Plot with Relative Risk

III. CLUSTER PLOT

A cluster plot is a technique for conducting point pattern analysis. It is a bivariate plot visualizing a partition or clustering of the data. It plots the data points and a representation of the located cluster.

We can use hierarchical clustering to run on a variety of similarity matrices based on pairwise similarity measures. Spearman 2 is selected for similarity measure. It measures the proportion of subjects having both AEs. The SAS procedure DISTANCE is used to produce a proximity matrix. The procedures CLUSTER and TREE are used to produce the hierarchical clustering tree. Figure 5 shows a sample AE cluster tree.

Figure 5. Sample AE Cluster Plot

5 IV. KAPLAN-MEIER TIME-TO-EVENT PLOT

Survival analysis is a class of statistical methods for studying the occurrence and timing of events. In a clinical trial setting, survival analysis is used to study the time-to-clinical events. Several procedures are provided from the SAS System for survival analysis. Procedure LIFETEST is used to produce a Kaplan-Meier time-to-event plot. Some SAS procedures produce ODS Tables as procedure output for further analysis.

Figure 6. Sample Cumulative Incidences of GI AE

This graph uses ODS output Tables from PROC LIFETEST as input for graphics presentation. The graphics are plotted on 4 panels, 3 panels for graphics while statistical test and censor information are shown in text table format on fourth panel. SAS 9.1 includes an experimental feature, ODS (ODS Graphics for short), and ODS Graphical Template Language (GTL). GTL is used to produce the graph. A template is created with a style to govern the graphics display.

The template consists of a 4-panel display, including:

1) a plot of estimated survival (Kaplan-Meier plot) with confidence bands, 2) a negative log of estimated survival, 3) a log of negative log of estimated survival, and 4) a table format of test of equality and censored summary statistics.

The panels of negative log plot and log of negative log plot are graphical methods for evaluation of model fit and graphical diagnostics. If the distribution of event times is exponential, a plot of negative log versus time should yield a straight line with an origin at 0. If event times have a Weibull distribution, a plot of log of negative log plot versus log time should also be a straight line. The text panel provides 2 tables of statistics, one for a test of equality over the strata, and the other for a censored summary table.

6

Figure 7. Time to Clinical Event Plot

Four-panel survival plots provide essential statistical testing information and a graphical presentation from the survival analysis.

SAS GTL is based on Java technology and provides a tool tip interactive feature. The tool tip feature allows the programmer to specify the information for display when the cursor is positioned over a graph element. Figure 8 illustrates a red arrow as cursor positioned over a graph element and a drop-down text box with graph element relevant information.

Figure 8. Data Tips Illustration

7 V. TREE DIAGRAM

A tree structure is a method of representing the hierarchical structure in a graphical form. It is named a “tree structure” because the graph looks like a tree. In graph theory, a tree is a collection of connected nodes. Every finite tree structure has a member that has no superior. This node is called the “root” node. The lines connecting nodes are called “branches” or “links”. Nodes without children are called “leaves” or “end-nodes”. A node is a “parent” of another node if it is one step higher in the hierarchy and closer to the root node.

Tree structures are used in many applications, such as hierarchical organizational structures, binary search tree, decision tree, partition tree, etc. There are many ways of visually representing tree structures. The most commonly used method is a classical node-link diagram, that connects nodes together with line segments. Figure 9 shows a tree structure from a Treeview applet with a root node in the center of the presentation. Figure 10 shows a recursive partition tree for AE occurrence prediction.

Figure 9. A Treeview of Serious AEs

Figure 10. A Recursive Partition Tree for AE Occurrence Prediction V. RADAR CHART

A radar or star chart graphically shows the size of the target numeric variable among categories. The procedure GRADAR is a new graphical procedure provided in SAS/GRAPH 9.1 for creating radar charts. A radar chart is a type of graphical presentation using spokes to illustrate the size of the target numeric variable among categories. 8 On a radar chart, the chart statistics are displayed along spokes from the center of the chart. The radar charts are called different names, because they comprise different types. These new types of data displays provide the novel look of graphical presentations, including interactive features.

The interactive features are:

1. By an ActiveX control – (pop-up data tips, drill-down links, and interactive menus) 2. By a Metaview applet – (data tips, drill-down links, or some interactivity such as zooming, panning, and slide shows). 3. By Static Images – (data tips, and drill-down links, or animation).

Figure 11 shows a radar chart for serious AEs display. Figure 12 shows the time trends in incidence of AEs.

Figure 11 Radar Chart for Serious AEs Display

Figure 12 Time Trends in Incidence of AEs VI. CONSTELLATION DIAGRAM

A constellation diagram is a representation of a modular scheme. It displays the data on two dimensional x-y plots with nodes and links. Links are used to describe relationships among component nodes. Different shapes and styles, with colors, are available for the user to select for the representation. Figure 13 is a sample

9 constellation diagram output from the DS2CONST macro. The Body System terms in the AE data provide the nodes and links and the AE sequence counts. The sizes and colors of nodes and links reflect the magnitude.

The Constellation Applet supports the following interactive features. It allows: 1) zooming/panning or fish-eye distortion on the portion of the diagram that is in the center of the display, 2) an embedded scroll bar to subset the links and nodes, 3) clicking and dragging the diagram to change the portion of the diagram, 4) a pop-up menu that has several functions, such as highlighting specific links and searching for specific nodes. 5) pop-up data tips for links and nodes. 6) hotspot links to Web sites or files.

One of the features from macro DS2CONST is to allow the user to assign the node’s location in x-y coordinates. This feature makes the DS2CONST a powerful tool for tree presentations.

Figure 13. Constellation Display of AE Onset Sequence

VII. NEURAL NETWORK DIAGRAM

An artificial neural network is a computer application that attempts to mimic the neurophysiology of the human brain by finding patterns of data in a representative sample. Neural network is a class of flexible models, discriminant models, and data reduction models, which are interconnected in a nonlinear dynamic system. This type of application can help an analyst make a prediction about clinical events for safety risk rate or clinical responder for efficacy prediction. Figure 14 shows a theoretical neural network diagram.

Figure 14. Theoretical Neural Network Diagram

10 Data Mining (DM) involves the use of software, statistical and graphical techniques to identify valid, non-trivial, previously unknown, interesting relationships/global patterns that exist in a large database. SAS EM provides a process flow diagram approach for DM tasks. The process flow diagram is a sequence of logical steps to build a DM project. Each step is defined by a visual flowchart depicting input, analyses and output. Neural Network is one of the DM techniques used for supervised prediction problems. The building blocks of an artificial neural network are the input layer, hidden layers, and the output layer. The input layer is composed of units that correspond to each input variable. The hidden layers are composed of hidden units. Each hidden unit outputs a non-linear function of a linear combination of its inputs – the activation function. The linear combination is the net output. The non-linear transformation is the activation function. The output layer has units corresponding to the target. With multiple target variables, there are multiple output units. The network diagram is a representation of an underlying . The unknown parameters (weights and biases) correspond to the connection between the units. The data used for network diagram are from SAS EM neural network modeling. The weight and link output during the modeling process is saved as an SAS dataset and read into a SAS/GRAPH constellation diagram program to generate the network diagram. The diagram is illustrated below.

Figure 15. Neural Network Diagram for AE Events Variables in the SAS dataset determine the size and color of nodes, as well as the width and color of the lines between nodes. The SAS Constellation diagram provides two interactive features: 1) a slider bar which allows a user to choose how many of the links on the diagram are displayed, and 2) data tips – which are activated when the viewer moves the mouse over a graphical element causing a text box to display extra information.

CONCLUSION

The reported AEs are the major components of clinical safety data. The clear and efficient graphical presentations can help in interpreting AE data and detecting safety signals. The graphical techniques and methods discussed in this paper could be effectively used to visualize AE data patterns, time trends, clusters, and data reduction, as well as to explore rare AE events.

The interactive features are very desirable and enable the user to: 1) change graph properties without rerunning the job, 2) include extra information in the output, and 3) control the subsetting nodes and links allowing the viewer to focus on stronger association nodes.

The SAS system procedures used in this paper are: CORR, DISTANCE, CLUSTER, TREE, GRADAR, GCHART, GPLOT. GREPLAY, DS2CONST macro, DS2TREE macro, and ODS Graphics. The computations and display power from the SAS System are fully utilized in this paper for the display of AE data and for the display layout design.

11 ACKNOWLEDGMENTS

The author wants to thank GSK Group Director Bob Schriver and GSK BDS Graphics Team members for their comments and suggestions.

REFERENCES

[1] Chuang-Stein C., Le, V., and Chen, W.(2001): Recent Advancement in the Analysis and Presentation of Safety Data, Drug Information Journal, Vol. 35, pp 377-397, 2001

[2] Friendly, M., (1991): SAS System for Statistical Graphics, SAS Institute, Cary NC.,USA

[3] Friendly, M., (2001): Visualizing Categorical Data, SAS Institute, Cary NC.,USA

[4] Harrell, E.F. and Burgan, T.M.(2001): Exploratory and Graphical Analysis of Clinical Data, Twenty-Fourth Annual Midwest Biopharmaceutical Statistics Workshop, Muncie, Indiana, 21-23 May, 2001

[5] Harrell, E.F.(2005): Exploratory Analysis of Clinical Safety Data to Detect Safety Signals , FDA Visiting Professor Lecture Series, WHITE Oak, MD, Oct. 11, 2005

[6] McQuitty, L. L, (1968): Multiple clusters, types, and dimensions from iterative intercolumnar correlational analysis, Multivariate Behavioral Research 3, pp 465-477.

[7] SAS Institute Inc.,(2003): SAS® OnlineDoc 9. , Cary NC. http://support.sas.com/91doc/docMainpage.jsp .

[8] SAS Institute Inc.,(2004): Base SAS® 9.1 Procedure Guide , Cary NC. SAS Institute Inc.

[9] SAS Institute Inc.,(2004): SAS/GRAPH® 9.1 Reference, Volumes 1 and 2 , Cary NC. SAS Institute Inc.

[10] SAS Institute Inc.,(2004): SAS/STAT® 9.1 User’s Guide , Cary NC. SAS Institute Inc.

[11] Wishart, D., (1999): ClustanGraphics3 Interactive Graphics for . Published in: Classification in the Information Age, Gaul, W., and Locarek-Junge, H., (Eds), Springer 1999, pp-268-275

[12] Williams, M. (2005): Signal Detection in Clinical Trials: An Industry , GSK, 2005

[13] Williams, M. (2006): The Problems with Safety Data: Tackling the Issues that Arise, DIA 18 th Annual Euro Meeting, Paris France, 2006

[14] Yeh, S.T.(2005): Using Interactive Treeview Diagrams for Clinical Adverse Experiences Records , Proceedings of the North East SAS Users Group 2005 Conference, Paper pos8 , September 2005.

[15] Yeh, S.T.(2005): Using RADAR Chart to Display Clinical Data , Proceedings of the North East SAS Users Group 2005 Conference, Paper pos5, September 2005.

[16] Yeh, S.T.(2005): Customizing ODS Statistical Graphics , Proceedings of the 30th SAS Users Group International Conference, Paper 183-30, April 2005.

[17] Yeh, S.T.(2006): Interactive Graphs from the SAS System , Proceedings of the 31th SAS Users Group International Conference, Paper 181-31, March 2006.

TRADEMARKS

SAS is a registered trademark of SAS Institute Inc., Cary, NC, USA in the USA, and other countries. ¡ indicates USA registration.

Author Shi-Tao Yeh, Ph. D. (610) 787-3856 (W) E-mail: [email protected]

12