Self-generating CONSORT diagram

library(consort)
data(dispos.data)

The goal of this package is to make it easy to create CONSORT diagrams for the transparent reporting of participant allocation in randomized, controlled clinical trials. This is done by creating a standardized disposition data, and using this data as the source for the creation a standard CONSORT diagram. Human effort by supplying text labels on the node can also be achieved. Below is the illustration of the CONSORT diagram creating process for the two different methods.

Prepare test data

In a clinical research, we have a participants disposition data. One column is the participants’ ID, and the following columns indicating the status of the participants at different stage of the study. One can easily derive the number of participants at different stage by counting the number of participants on-study excluding the participants who are excluded.

Description

This function developed to populate consort diagram automatically. But to do so, a population disposition data should be prepared. The following data is prepared for demonstration provided in the package. The data can be loaded with data(dispos.data), the variables in this table is explained below:

#>   trialno exclusion1 induction exclusion2 exclusion  arm  arm3 sbujid_dosed
#> 1    1086       <NA>      1086       <NA>      <NA> Conc Trt C         1086
#> 2    1418       <NA>      1418       <NA>      <NA>  Seq Trt B         1418
#> 3    1502       <NA>      1502       <NA>      <NA>  Seq Trt C         1502
#> 4    1846       <NA>      1846       <NA>      <NA>  Seq Trt A         1846
#> 5    1303       <NA>      1303       <NA>      <NA>  Seq Trt A         1303
#> 6    1838       <NA>      1838       <NA>      <NA> Conc Trt B         1838
#>   subjid_notdosed followup lost_followup assessed           no_value mitt
#> 1            <NA>     1086          <NA>     1086               <NA> 1086
#> 2            <NA>     1418          <NA>     1418               <NA> 1418
#> 3            <NA>     1502          <NA>     1502 Protocol deviation 1502
#> 4            <NA>     1846         Death       NA               <NA> 1846
#> 5            <NA>     1303      Withdraw       NA               <NA> 1303
#> 6            <NA>     1838          <NA>     1838               <NA> 1838

Usage

Basic logic:

  1. The vertical node are the number of patients in the current node, no dropout reasons of inclusion reasons should be provided.
  2. Side box is only used for the drop outs. It include information about the number of patients excluded and reasons.
  3. Any participants with exclusion reasons will not be included in the next vertical node box after excluded. So the subject id can be used multiple times to indicate how many patients left in the current node.
  4. The node labels, for example visit number or phase, can only horizontally align to a vertical main nodes, not an exclusion box.
  5. If more than 2 treatment allocation is present, all the exclusion box after the allocation will be aligned to the right

At each stage, number of non-missing values will be counted. Set kickoff_sidebox=TRUE to remove the observations counted in the side box in the remaining nodes.

Self-generating function

To generate consort diagram with data.frame, one should prepare a disposition data.frame.

consort_plot(data,
             orders,
             side_box,
             allocation = NULL,
             labels = NULL,
             cex = 0.8,
             text_width = NULL,
             widths = c(0.1, 0.9))

Manual

Functions are mainly in three categories, main box, side box and label box. Others include building function. These are the functions used by the self generating function. These box functions require the previous node and text label. - add_box: add main box, no previous nodes should be provided if this is the first node. - add_side_box: add exclusion box. - add_split: add allocation box, all nodes will be split into groups. The label text for this node and following nodes should be a vector with a length larger than 1. - add_label_box: add visiting or phasing label given a reference node. - build_grid: you don’t need this unless you want the grob object. - build_grviz: you don’t need this unless you want the Graphviz code.

Plotting engine

The package supports two plotting engines: grid (grid package) and Graphviz. The default was grid, plot(g) (assuming g is the consort plot object) will use the grid to draw the plot. The coordinates are calculated internally, one can simply use build_grviz(g) to get a built grob object. This means you can do use patchwork package to combine different plots. You can use the example below to add title and footnotes.

library(patchwork)

wrap_elements(build_grid(g)) + plot_annotation(
  title = 'Consort diagram',
  subtitle = 'Flow chart of the XX study',
  caption = 'Disclaimer: None of these plots are insightful'
)

Although many efforts have been made to draw the plot as easy as possible, you may still find some plots are difficult to draw. You may not be happy about the final output or want the plot in a svg format to be used in the websites. You can simply plot(g, grViz = TRUE) to draw the plot with Graphviz. You need to install DiagrammeR package to do so. If you want to do some twicking to the plot, you can use build_grviz(g) to extract the dot file and edit. You can use Rstudio to render the plot or other tools you prefer. Use the example below to dump the dot file.

cat(build_grviz(g), file = "consort.gv")

Working example (self generation)

Single arm

out <- consort_plot(data = dispos.data,
                    orders = c(trialno = "Population",
                               exclusion = "Excluded",
                               trialno = "Allocated",
                               subjid_notdosed = "Not dosed",
                               followup = "Followup",
                               lost_followup = "Not evaluable for the final analysis",
                               mitt = "Final Analysis"),
                    side_box = c("exclusion", "subjid_notdosed", "lost_followup"),
                    cex = 0.9)
plot(out)

Multiple phase and multiple arms

g <- consort_plot(data = dispos.data,
                  orders = c(trialno = "Population",
                             exclusion1    = "Excluded",
                             induction   = "Induction",
                             exclusion2    = "Excluded",
                             arm3     = "Randomized patient",
                             subjid_notdosed = "Not dosed",
                             mitt = "Final miTT Analysis"),
                  side_box = c("exclusion1", "exclusion2", "subjid_notdosed"),
                  allocation = "arm3",
                  labels = c("1" = "Screening", "2" = "Month 4",
                             "3" = "Randomization", 
                             "5" = "End of study"),
                  cex = 0.7)
plot(g)

Multiple stratification with node stack

In some cases, two level randomisation/stratification is desired. A typical case is the factorial design. Simply provide two allocation variables, but more than two splits are not supported.

Here is a simple example on how to do it.

df <- dispos.data[!dispos.data$arm3 %in% "Trt C",]
g <- consort_plot(data = df,
                  orders = c(trialno = "Population",
                             exclusion1    = "Excluded",
                             induction   = "Induction",
                             exclusion2    = "Excluded",
                             arm = "Randomized patient",
                             arm3     = "",
                             subjid_notdosed = "Not dosed",
                             mitt = "Final miTT Analysis"),
                  side_box = c("exclusion1", "exclusion2", "subjid_notdosed"),
                  allocation = c("arm", "arm3"),
                  labels = c("1" = "Screening", "2" = "Month 4",
                             "3" = "Randomization", 
                             "6" = "End of study"),
                  cex = 0.7)
plot(g)

In a more general cases, one may want to keep the count inside the vertical box instead of excluding. In this cases, one can provide multiple variables in a list. The first variable will simply be counted, the remaining variables will be itemised.

# We will exclude one arm to avoid too many arms.
df <- dispos.data[!dispos.data$arm3 %in% "Trt C",]
p <- consort_plot(data = df,
                  orders = list(c(trialno = "Population"),
                                c(exclusion = "Excluded"),
                                c(arm     = "Randomized patient"),
                                # The following two variables will be stacked together
                                c(arm3     = "", # Should not provide a value to show the actual arm
                                  subjid_notdosed="Participants not treated"),
                                # The following two variables will be stacked together
                                c(followup    = "Pariticpants planned for follow-up",
                                  lost_followup = "Reason for tot followed"),
                                c(assessed = "Assessed for final outcome"),
                                c(no_value = "Reason for not assessed"),
                                c(mitt = "Included in the mITT analysis")),
                  side_box = c("exclusion", "no_value"), 
                  allocation = c("arm", "arm3"), # Two level randomisation
                  kickoff_sidebox = FALSE,
                  labels = c("1" = "Screening", "2" = "Randomization",
                             "5" = "Follow-up", "7" = "Final analysis"),
                  cex = 0.7)

plot(p)

Working example (human effort)

The previous is to easily generate a consort diagram based on a disposition data, here we show how to create a consort diagram by providing the label text manually.

Provide text

library(grid)
# Might want to change some settings
options(txt_gp = gpar(cex = 0.8))

txt0 <- c("Study 1 (n=160)", "Study 2 (n=140)")
txt1 <- "Population (n=300)"
txt1_side <- "Excluded (n=15):\n\u2022 MRI not collected (n=3)\n\u2022 Tissues not collected (n=4)\n\u2022 Other (n=8)"

# supports pipeline operator
g <- add_box(txt = txt0) |>
  add_box(txt = txt1) |>
  add_side_box(txt = txt1_side) |> 
  add_box(txt = "Randomized (n=200)") |> 
  add_split(txt = c("Arm A (n=100)", "Arm B (n=100)")) |> 
  add_side_box(txt = c("Excluded (n=15):\n\u2022 MRI not collected (n=3)\n\u2022 Tissues not collected (n=4)\n\u2022 Other (n=8)",
                       "Excluded (n=7):\n\u2022 MRI not collected (n=3)\n\u2022 Tissues not collected (n=4)")) |> 
  add_box(txt = c("Final analysis (n=85)", "Final analysis (n=93)")) |> 
  add_label_box(txt = c("1" = "Screening",
                        "3" = "Randomized",
                        "4" = "Final analysis"))
plot(g)

Missing nodes and multiple split

There might be some cases that the nodes will be missing, this can be handled as well. You can also have multiple splits, but multiple splits for the grid hasn’t been implemented yet. You can use plot(g, grViz = TRUE) to produce the consort plot.

g <- add_box(txt = c("Study 1 (n=8)", "Study 2 (n=12)", "Study 3 (n=12)"))
g <- add_box(g, txt = "Included All (n=20)")
g <- add_side_box(g, txt = "Excluded (n=7):\n\u2022 MRI not collected (n=3)")
g <- add_box(g, txt = "Randomised")
g <- add_split(g, txt = c("Arm A (n=143)", "Arm B (n=142)"))
g <- add_box(g, txt = c("", "From Arm B"))
g <- add_box(g, txt = "Combine all")
# Length needs to be the same as previous one, use a list here.
g <- add_split(g, txt = list(c("Process 1 (n=140)", "Process 2 (n=140)",
                               "Process 3 (n=142)")))

plot(g, grViz = TRUE)

Two-level stratification

Some studies may first stratify patients then randomise, a factorial design for example. This will require two-level stratification. One can simply provide a list of the same length of the previous node after a split. More than two splits are not supported.

g <- add_box(txt = "Patient consented (n=200)")
g <- add_side_box(g, txt = "Excluded (n=10):\n\u2022 MRI not collected (n=3)\n\u2022 Other (n=7)")
g <- add_box(g, txt = "Randomised (n=190)")
g <- add_split(g, txt = c("Randomised study (n=144)", "Preference study (n=146)"))

# Provide a list here
g <- add_split(g, txt = list(c("Allocated to surgery (n=70)", 
                                "Allocated to medicine (n=74)"),
                             c("Allocated to surgery (n=47)", 
                              "Allocated to medicine (n=48)",
                               "Placebo (n=51)")))

g <- add_side_box(g, txt = c("Excluded (n=3):\n\u2022 Withdrawn before surgery (n=2)\n\u2022 Declined (n=1)", 
                             "", 
                             "Excluded (n=1):\n\u2022 Withdrawn before surgery (n=1)",
                             "Excluded (n=3):\n\u2022 Withdrawn before surgery (n=1)\n\u2022 Declined (n=1)",
                             "Excluded (n=10):6\n\u2022 Declined (n=10)"),
                  side = rep("right", 5))
g <- add_box(g, txt = c("Analysed (n=67)", "Analysed (n=74)",
                        "Analysed (n=46)", "Analysed (n=45)", "Analysed (n=41)"))

plot(g)

Using disposition table

options(txt_gp = gpar(cex = 0.8))

dispos.data$arm <- factor(dispos.data$arm)

txt <- gen_text(dispos.data$trialno, label = "Patient consented")
g <- add_box(txt = txt)

txt <- gen_text(dispos.data$exclusion, label = "Excluded", bullet = TRUE)
g <- add_side_box(g, txt = txt)   

g <- add_box(g, txt = gen_text(dispos.data$arm, label = "Patients randomised")) 

txt <- gen_text(dispos.data$arm)
g <- add_split(g, txt = txt)

# Exclude subjects
dispos.data <- dispos.data[is.na(dispos.data$exclusion), ]

txt <- gen_text(split(dispos.data[,c("subjid_notdosed")], dispos.data$arm),
                label = "Not dosed", bullet = TRUE)
g <- add_box(g, txt = txt, just = "left")

txt <- gen_text(split(dispos.data$mitt, dispos.data$arm),
                label = "Primary mITT analysis")
g <- add_box(g, txt = txt)

g <- add_label_box(g, txt = c("1" = "Baseline",
                              "5" = "Final analysis"))

plot(g)

For Shiny and HTML

Although all the efforts has been made to precisely calculate the coordinates of the nodes, it is not very accurate due to limit of my own knowledge. But you can utilize the DiagrammeR to produce plots for Shiny and HTML by setting grViz = TRUE in plot. You can get Graphviz code with build_grviz of the plot. In addition, use DiagrammeRsvg and rsvg save plot in various formats.

plot(g, grViz = TRUE)

Use with Quarto

Quarto has native support for embedding Graphviz diagrams. You can plot the flowchart without any printing method.

```{r}
cat(build_grviz(g), file = "consort.gv")
```

```{dot}
//| label: consort-diagram
//| fig-cap: "CONSORT diagram of study XXX"
//| file: consort.gv
```

Saving plot

In order to export the plot to fit a page properly, you need to provide the width and height of the output plot. You might need to try different width and height to get a satisfying plot.You can use R basic device destination for the output. Below is how to save a plot in png format:

# save plots
png("consort_diagram.png", width = 29, 
    height = 21, res = 300, units = "cm", type = "cairo") 
plot(g)
dev.off()

Or you can use ggplot2::ggsave function to save the plot object:

ggplot2::ggsave("consort_diagram.pdf", plot = build_grid(g))

Or save with DiagrammeRsvg and rsvg to png or pdf

plot(g, grViz = TRUE) |> 
    DiagrammeRsvg::export_svg() |> 
    charToRaw() |> 
    rsvg::rsvg_pdf("svg_graph.pdf")