 Γεια σας! Ευχαριστώ για την παρακολουθή σας! My name is Giorgio Cetidis, I am a PhD candidate at the University of Ioannina and in this talk I will present graphical tools for visualizing the results of network meta-analysis of multi-component interventions. The outline of this presentation is the following. Firstly, I will make a brief overview of network meta-analysis and I will present the objectives of this work. Secondly, I will introduce the viscom arc parkas and present the motivational example which will be used to demonstrate the applicability of the proposed visualization tools. Then I will guide you through them and I will close by presenting the conclusions of this work. Network meta-analysis is an established statistical method for synthesizing evidence from studies that compare multiple interventions. Networks usually include interventions that consist of multiple and possible iterative components such as the characteristics of the intervention, the mode of delivery if it is delivered face-to-face or remotely, individually or in groups, the type of provider, the location etc. These interventions are referred in the literature as multi-component or complex. When they are present, the interest usually lies on the identification of the best component. The network meta-analysis identifies the best interventions in terms of efficacy and as a result we can identify the most efficacious combination of components. In practice, we may observe components that appear in the most efficacious interventions but also in the least efficacious interventions. In such cases, it is hard to conclude about the efficacy of the other line components. Also, in practice it is likely to have a sparse network with many nodes and few head-to-head studies. In such networks, nodes are usually formed by studies involving these nodes and not by indirect evidence. Therefore, it is quite likely to have network meta-analysis estimates that are mainly driven by the study estimates and torsos that are confounded by the study characteristics. These issues may challenge the transitivity assumption. To identify the most promising components, the component meta-analysis is commonly used. However, component meta-analysis assumes transitivity for the component's effects in multi-component interventions. This is a strong assumption and sometimes is hard to defend. It is even harder when the network is sparse and this is because the summary estimates of meta-analysis are mainly informed by direct evidence and are prone to confounding. The objective of this work was to develop novel ways of identifying the most promising components and exploring their behavior. This was done by associating the presence or the absence of components with efficacy or effectiveness. The implementation of the proposed tools can be performed through the R-Packets viscom. Viscom is a friendly user R-Packets offering a powerful toolkit on exploring the behavior of the components in a network meta-analysis of multi-component interventions. It was recently published on C-RAN and it is also available on GitHub. The input that is required in order to use the packets is a network meta-analysis model as obtained from the R-Packets net method. Viscom can be used only in connected networks than the group multi-component interventions. Also, the reference groups will be defined in the network meta-analysis model as it is required from the majority of the functions. The packets by default uses the random effects network meta-analysis model and the network meta-analysis delta effect estimates. In total, eight functions are included in the package, which is the COM desk for component descriptions analysis, the COM graph for the components network graph, LOCOS for the living one-component combination outscout plot, the WATER COM for the waterfall plot, the HEAT COM for the components heat plot, the SPEC for the specific component combination violent plot, the DENSE COM for the components density plot, and the RANKY PLOT for the components ranky plot. To demonstrate the uses of the packets functions, we will use as an example a network that evaluates self-mask interventions on reducing the body mass index. Interventions are masked, as this result will be submitted soon to a medical journal and cannot be revealed beforehand. The network includes in total 135 studies, 56 nodes and 11 components. As you can see from the network plot, most comparisons involve node A, which is a control group and it will be used as a reference in the network analysis model. Also, in the network there are many nodes, many comparisons, but few heat-to-heat studies. Therefore, the assumptions of network network meta-analysis and component network meta-analysis will be challenged due to the sparsity of the network. Exploring the geometry of the network is essential, especially when you are dealing with large networks with complex structures. Component descriptive analysis can assist in achieving this goal. Through them, facts and conducts we can perform a descriptive analysis regarding the frequency of the components and identify the most frequent component combinations. The only input that it is required from the function is a network meta-analysis model, which is denoted by the argument model. The rest arguments refer to the operator symbol between components. If the heat matrix could be plotted, the hugus for the percentages is then the fraction values and the number of digits. As an output, the function returns three objects, the cross-tables, the frequency table and the heat matrix. The first object is a cross-table, which describes the number of studies where the corresponding component combination was observed. The second object is a table that displays the component's frequency and provides information concerning the number of studies in which the underlying component is included in all arms, in at least one arm and not included in any arm. The third object is the heat matrix, which is a color visualization of the cross-table. Diagonal elements refer to components and in parenthesis the proportion of the study arms that include the underlying component is presented. Of diagonal elements, refer to the combination of components and in parenthesis the proportion of the study arms with both components out of those study arms that include the component in the row is presented. For example, the element that corresponds to column E and row A indicates that 42 study arms include these components, but also that component E was always included in the interventions that include component H. To identify more easily the most frequent combinations, the table is colored based on the relative frequencies. The more it turns the color, the larger the percentage in the parenthesis. Therefore, dark red colors indicate large percentages. From the figure we can see that the most dark red colors were observed in the column E. This indicates that the component E was almost always part of the intervention. By stating the argument percentage equal to false, the heat matrix displays fraction of numbers instead of percentages. Another tool for exploring the network geometry is the component's network graph, which visualizes the frequency of component's combinations and identifies the most frequent. It can be constructed by using the function from graph, which by default plots the five most common combinations. The nodes here represent the individual components found in the network, while the edges represent the combination of components. Each edge's color represents one of the unique combinations of components found in the network. Also, the thickness is proportional to the number of arms in which the corresponding combination was observed. In this example, the component's network graph indicates that the most frequent component is the component A, followed by the combination between E and C. Note also that components A and D were not combined with any other component, as expected, since they refer to control groups. The two control groups can be excluded through the argument EXEL. Also, the component's network graph is adjusted to display the six most frequent component combinations by stating accordingly the argument must F. The leaving one component combination outscatter plot can be used to explore whether the inclusion or exclusion of a set of components has a positive or negative impact on the outcome. Also, it can be used to evaluate visually the activity assumption and it can be constructed by using the function Locos. The function by default uses the relative effect to estimates and visualizes the set of interventions that differ by one component. Dots close to the line of equality signify no impact on the outcome. Dots above this line indicate that the enemy effect estimates are larger when the component is not included in the intervention. Therefore, for beneficial outcomes in which small values are considered as bad values, dots above the line indicate that the inclusion of a component hampers the intervention effect, while dots below this line signify a component that decreases efficacy. The opposite holds for harmful outcomes. Another feature of the scatter plot is that we can visually evaluate the activity assumption. Additivity implies that the inclusion or exclusion of a component has the same impact on interventions that differ by one specific component. This is expressed in the scatter plot by a line parallel to the line of equality. By properly setting the argument combination, we can define the component of interest. For example, by setting the argument combination equal to C, we can visualize the set of interventions that differ by the component C. Here, most dots are below the line of equality, indicating that the inclusion of component C hampers the intervention effect. Also, we do not observe any line parallel to the line of equality, indicating that the additivity assumption in the CNA model may not hold. The histogram in the axis describes the distribution of the dots. The color of the histogram can be changed through the argument histogram color. Also, by setting the argument histogram equal to false, histograms are not displayed. The function can be easily extended to component combinations through the argument combination. If, for example, we are interested in the combination between components B and C, we can visualize the set of interventions that differ by these components by setting the argument combination equal to B plus C. It should be noted that the spacing among the components does not affect the results of this comb. If, for example, we delete the space after the separator symbol, the results will be the same. Lastly, we can use z values instead of relative effects. By setting the argument z value equal to true, an alternative to the live one component combination out scatter plot is a waterfall plot which evaluates also whether the inclusion of a set of components has a positive or negative impact on the outcome. The waterfall plot can be constructed through the function watercomb. The function by default uses relative effect estimates and visualizes the set of interventions that differ by one component. By setting the argument combination equal to C, we visualize all the set of interventions that differ by the component C. In this example, which is considered as a harmful outcome, bars above zero indicate that the inclusion of components C in hamper's efficacy while below zero that increases efficacy. From the plot, we see that most bars are above zero, indicating that the inclusion of components C may reduce the intervention efficacy. The function can be easily extended to components combinations through the argument combination. If, for example, we are interested in the combination between components B and C, we can visualize the set of interventions that differ by these components by setting the argument combination equal to B plus C. Lastly, we can use z values instead of relative effects by setting the argument z value equal to true. Components he plot visualizes the efficacy of the two-by-two component combinations and it can be constructed through the function hit-com. Each element summarizes the enemy estimate where the corresponding component combination was observed. As a summary measure, the media nor the mean can be used. The number of nodes that include the corresponding component combination is also provided in the parentheses of itself. Letter x is used to highlight any combination of components, but are not observed in the network. To identify more easily the most promising components, each element was colored according to the magnitude of its effect. Green color is used to reflect an effect estimate on the desired direction, while red color is used to reflect an effect estimate on the opposite direction. The outcome scenario, harmful or beneficial, is obtained from the network analysis model. The ascertainty around the estimates is also reflected by the size of the gray boxes. The smaller the box, the more confident we are about to find. By setting the argument media equal to false, the mean is used as a summary measure instead of the media. Also, by setting the argument z value equal to true, z values are used instead of relative effects. Lastly, the number of nodes that is displayed in the parentheses of each element can be excluded by setting the argument 3 equal to false. Similar to the component's heat plot, is a specific component combination violin plot, which also visualizes the efficacy of the components. It produces violins based on the interventions and includes the component combination of interest, and also it can be extended to the number of components. The violin plots can be easily constructed by using the function spec. The function by default visualizes the distribution of the relative effects for each component. Through the function combination, we can define the combination of interest. If, for example, we are interested in the components b, e, and the combination between them, by setting the argument appropriately, we obtain three violins. One for the interventions that include component b, one for the component e, and one for the interventions that include both components. Note also that the size of the dots, in the case where relative effects are used, is by default proportional to the precision of the estimate. Function can be easily adjusted to component's numbers by setting the argument component's numbers equal to true. The violins now are constructed based on the number of components that are included in the interventions. Therefore, we can explore if the number of components affect the intervention's efficacy. The number of components can be also grouped by using the argument group. Here, we have grouped the violins based on the interventions that include one component, one to three components, four to five components, and six or more components. The violins can be easily adjusted to display z-values instead of relative effects by setting the argument z-value equals true. Another way to visualize the efficacy of the components is by using density plots. Component's density plots can be easily constructed through the function density. The function requires to define the component of interest. Let us assume that component B is the component of interest. By setting the argument combination equals to B, the function produces two density plots. One, based on the interventions that include component B, and the other, based on the interventions that do not include component B. The function can be easily extended to component's combination by setting properly the argument combination. Here, we have constructed a density plot for the combination between components C, B, and I. The argument combination includes more than one element. The number of densities is equal with the length of the argument. And each density is based on the intervention that includes the relative component combination. For example, here the argument combination includes three elements. Therefore, three densities will be displayed based on the interventions that include components E and B, component E, and component C. Also, by setting the argument by only equal to true, the densities are visualized through the violins. Lastly, we can use z values instead of relative effects. By setting the argument z value equal to true. In the case where multiple outcomes are included in our analysis, we can identify the best component by using the component's ranking plot. We can easily construct it through the function ranking plot. Here, we have constructed the ranking plot based on three different outcomes. The high density lipoprotein cholesterol, the waist size, and the body mass index. The function requires as input a list that contains analysis models. In the ranking plot, each sector summarizes the pitch scores from the interventions that include the corresponding components. The calculation of the pitch scores is based on the outcome nature, beneficial or harmful, which is defined in the network analysis model. Also, to identify more easily the best components, the sectors are colored according to the ranking. Dark green color is used to reflect percentages close to 100, while dark red colors are used for percentages close to 0. As a summary measure, the median is used by default, but it can also be adjusted to use the mean instead. To conclude, these comps should be used when multi-component interventions are present in the network. The highlight of the package is that it offers a friendly user toolkit in order to explore the geometry of the network and the component's behavior. Also, it can provide viable insight on identifying the most promising components and it should be used in order to strengthen the network meta-analysis and component network meta-analysis. So, that's it. Thank you for your attention. I'm happy to answer any questions or comments and of course any ideas or suggestions are model welcome. Bye-bye.