AMR/vignettes/AMR.Rmd

---
title: "How to conduct AMR analysis"
author: "Matthijs S. Berends"
date: '`r format(Sys.Date(), "%d %B %Y")`'
output: 
  rmarkdown::html_vignette:
    toc: true
    toc_depth: 3
vignette: >
  %\VignetteIndexEntry{How to conduct AMR analysis}
  %\VignetteEncoding{UTF-8}
  %\VignetteEngine{knitr::rmarkdown}
editor_options: 
  chunk_output_type: console
---

```{r setup, include = FALSE, results = 'markup'}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 7.5,
  fig.height = 5
)
```

**Note:** values on this page will change with every website update since they are based on randomly created values and the page was written in [R Markdown](https://rmarkdown.rstudio.com/). However, the methodology remains unchanged. This page was generated on `r format(Sys.Date(), "%d %B %Y")`.

# Introduction

For this tutorial, we will create fake demonstration data to work with. 

You can skip to [Cleaning the data](#cleaning-the-data) if you already have your own data ready. If you start your analysis, try to make the structure of your data generally look like this:

```{r example table, echo = FALSE, results = 'asis'}
knitr::kable(dplyr::tibble(date = Sys.Date(),
                           patient_id = c("abcd", "abcd", "efgh"),
                           mo = "Escherichia coli", 
                           amox = c("S", "S", "R"),
                           cipr = c("S", "R", "S")), 
             align = "c")
``` 

## Needed R packages
As with many uses in R, we need some additional packages for AMR analysis. Our package works closely together with the [tidyverse packages](https://www.tidyverse.org) [`dplyr`](https://dplyr.tidyverse.org/) and [`ggplot2`](https://ggplot2.tidyverse.org) by [Dr Hadley Wickham](https://www.linkedin.com/in/hadleywickham/). The tidyverse tremendously improves the way we conduct data science - it allows for a very natural way of writing syntaxes and creating beautiful plots in R.

Our `AMR` package depends on these packages and even extends their use and functions.

```{r lib packages, message = FALSE}
library(dplyr)
library(ggplot2)
library(AMR)

# (if not yet installed, install with:)
# install.packages(c("tidyverse", "AMR"))
```

# Creation of data
We will create some fake example data to use for analysis. For antimicrobial resistance analysis, we need at least: a patient ID, name or code of a microorganism, a date and antimicrobial results (an antibiogram). It could also include a specimen type (e.g. to filter on blood or urine), the ward type (e.g. to filter on ICUs). 

With additional columns (like a hospital name, the patients gender of even [well-defined] clinical properties) you can do a comparative analysis, as this tutorial will demonstrate too.

## Patients
To start with patients, we need a unique list of patients. 

```{r create patients}
patients <- unlist(lapply(LETTERS, paste0, 1:10))
```

The `LETTERS` object is available in R - it's a vector with 26 characters: `A` to `Z`. The `patients` object we just created is now a vector of length `r length(patients)`, with values (patient IDs) varying from ``r patients[1]`` to ``r patients[length(patients)]``. Now we we also set the gender of our patients, by putting the ID and the gender in a table:

```{r create gender}
patients_table <- data.frame(patient_id = patients,
                             gender = c(rep("M", 135),
                                        rep("F", 125)))
```

The first 135 patient IDs are now male, the other 125 are female.

## Dates
Let's pretend that our data consists of blood cultures isolates from 1 January 2010 until 1 January 2018. 

```{r create dates}
dates <- seq(as.Date("2010-01-01"), as.Date("2018-01-01"), by = "day")
```

This `dates` object now contains all days in our date range.

#### Microorganisms
For this tutorial, we will uses four different microorganisms: *Escherichia coli*, *Staphylococcus aureus*, *Streptococcus pneumoniae*, and *Klebsiella pneumoniae*:

```{r mo}
bacteria <- c("Escherichia coli", "Staphylococcus aureus",
              "Streptococcus pneumoniae", "Klebsiella pneumoniae")
```

## Other variables
For completeness, we can also add the hospital where the patients was admitted and we need to define valid antibmicrobial results for our randomisation:

```{r create other}
hospitals <- c("Hospital A", "Hospital B", "Hospital C", "Hospital D")
ab_interpretations <- c("S", "I", "R")
```

## Put everything together

Using the `sample()` function, we can randomly select items from all objects we defined earlier. To let our fake data reflect reality a bit, we will also approximately define the probabilities of bacteria and the antibiotic results with the `prob` parameter.

```{r merge data}
sample_size <- 20000
data <- data.frame(date = sample(dates, size = sample_size, replace = TRUE),
                   patient_id = sample(patients, size = sample_size, replace = TRUE),
                   hospital = sample(hospitals, size = sample_size, replace = TRUE,
                                     prob = c(0.30, 0.35, 0.15, 0.20)),
                   bacteria = sample(bacteria, size = sample_size, replace = TRUE,
                                     prob = c(0.50, 0.25, 0.15, 0.10)),
                   amox = sample(ab_interpretations, size = sample_size, replace = TRUE,
                                 prob = c(0.60, 0.05, 0.35)),
                   amcl = sample(ab_interpretations, size = sample_size, replace = TRUE,
                                 prob = c(0.75, 0.10, 0.15)),
                   cipr = sample(ab_interpretations, size = sample_size, replace = TRUE,
                                 prob = c(0.80, 0.00, 0.20)),
                   gent = sample(ab_interpretations, size = sample_size, replace = TRUE,
                                 prob = c(0.92, 0.00, 0.08))
                   )
```

Using the `left_join()` function from the `dplyr` package, we can 'map' the gender to the patient ID using the `patients_table` object we created earlier:

```{r merge data 2, message = FALSE, warning = FALSE}
data <- data %>% left_join(patients_table)
```

The resulting data set contains 5,000 blood culture isolates. With the `head()` function we can preview the first 6 values of this data set:

```{r preview data set 1, eval = FALSE}
head(data)
```

```{r preview data set 2, echo = FALSE, results = 'asis'}
knitr::kable(head(data), align = "c")
```

Now, let's start the cleaning and the analysis!

# Cleaning the data
Use the frequency table function `freq()` to look specifically for unique values in any variable. For example, for the `gender` variable:

```{r freq gender 1, eval = FALSE}
data %>% freq(gender) # this would be the same: freq(data$gender)
```

```{r freq gender 2, echo = FALSE, results = 'markup'}
data %>% freq(gender, markdown = FALSE, header = TRUE)
```

So, we can draw at least two conclusions immediately. From a data scientist perspective, the data looks clean: only values `M` and `F`. From a researcher perspective: there are slightly more men. Nothing we didn't already know.

The data is already quite clean, but we still need to transform some variables. The `bacteria` column now consists of text, and we want to add more variables based on microbial IDs later on. So, we will transform this column to valid IDs. The `mutate()` function of the `dplyr` package makes this really easy:

```{r transform mo 1}
data <- data %>%
  mutate(bacteria = as.mo(bacteria))
```

We also want to transform the antibiotics, because in real life data we don't know if they are really clean. The `as.rsi()` function ensures reliability and reproducibility in these kind of variables. The `mutate_at()` will run the `as.rsi()` function on defined variables:

```{r transform abx}
data <- data %>%
  mutate_at(vars(amox:gent), as.rsi)
```

Finally, we will apply [EUCAST rules](http://www.eucast.org/expert_rules_and_intrinsic_resistance/) on our antimicrobial results. In Europe, most medical microbiological laboratories already apply these rules. Our package features their latest insights on intrinsic resistance and exceptional phenotypes. Moreover, the `eucast_rules()` function can also apply additional rules, like forcing <help title="ATC: J01CA01">ampicillin</help> = R when <help title="ATC: J01CR02">amoxicillin/clavulanic acid</help> = R.

Because the amoxicillin (column `amox`) and amoxicillin/clavulanic acid (column `amcl`) in our data were generated randomly, some rows will undoubtedly contain amox = S and amcl = R, which is technically impossible. The `eucast_rules()` fixes this:

```{r eucast, warning = FALSE, message = FALSE}
data <- eucast_rules(data, col_mo = "bacteria")
```

# Adding new variables
Now that we have the microbial ID, we can add some taxonomic properties:

```{r new taxo}
data <- data %>% 
  mutate(gramstain = mo_gramstain(bacteria),
         genus = mo_genus(bacteria),
         species = mo_species(bacteria))
```

## First isolates
We also need to know which isolates we can *actually* use for analysis.

To conduct an analysis of antimicrobial resistance, you must [only include the first isolate of every patient per episode](https://www.ncbi.nlm.nih.gov/pubmed/17304462) (Hindler *et al.*, Clin Infect Dis. 2007). If you would not do this, you could easily get an overestimate or underestimate of the resistance of an antibiotic. Imagine that a patient was admitted with an MRSA and that it was found in 5 different blood cultures the following weeks (yes, some countries like the Netherlands have these blood drawing policies). The resistance percentage of oxacillin of all \emph{S. aureus} isolates would be overestimated, because you included this MRSA more than once. It would clearly be [selection bias](https://en.wikipedia.org/wiki/Selection_bias).

The Clinical and Laboratory Standards Institute (CLSI) appoints this as follows:

> *(...) When preparing a cumulative antibiogram to guide clinical decisions about empirical antimicrobial therapy of initial infections, **only the first isolate of a given species per patient, per analysis period (eg, one year) should be included, irrespective of body site, antimicrobial susceptibility profile, or other phenotypical characteristics (eg, biotype)**. The first isolate is easily identified, and cumulative antimicrobial susceptibility test data prepared using the first isolate are generally comparable to cumulative antimicrobial susceptibility test data calculated by other methods, providing duplicate isolates are excluded.*
<br>[M39-A4 Analysis and Presentation of Cumulative Antimicrobial Susceptibility Test Data, 4th Edition. CLSI, 2014. Chapter 6.4](https://clsi.org/standards/products/microbiology/documents/m39/)

This `AMR` package includes this methodology with the `first_isolate()` function. It adopts the episode of a year (can be changed by user) and it starts counting days after every selected isolate. This new variable can easily be added to our data:
```{r 1st isolate}
data <- data %>% 
  mutate(first = first_isolate(.))
```

So only `r AMR:::percent(sum(data$first) / nrow(data))` is suitable for resistance analysis! We can now filter on it with the `filter()` function, also from the `dplyr` package:

```{r 1st isolate filter}
data_1st <- data %>% 
  filter(first == TRUE)
```

For future use, the above two syntaxes can be shortened with the `filter_first_isolate()` function:

```{r 1st isolate filter 2, eval = FALSE}
data_1st <- data %>% 
  filter_first_isolate()
```

## First *weighted* isolates
We made a slight twist to the CLSI algorithm, to take into account the antimicrobial susceptibility profile. Imagine this data, sorted on date:

```{r, echo = FALSE, message = FALSE, warning = FALSE, results = 'asis'}
weighted_df <- data %>%
  filter(bacteria == as.mo("E. coli")) %>% 
  # only most prevalent patient
  filter(patient_id == top_freq(freq(., patient_id), 1)[1]) %>% 
  arrange(date) %>%
  select(date, patient_id, bacteria, amox:gent, first) %>% 
  # maximum of 10 rows
  .[1:min(10, nrow(.)),] %>% 
  mutate(isolate = row_number()) %>% 
  select(isolate, everything())

weighted_df %>% 
  knitr::kable(align = "c")
```

Only `r sum(weighted_df$first)` isolates are marked as 'first' according to CLSI guideline. But when reviewing the antibiogram, it is obvious that some isolates are absolutely different strains and should be included too. This is why we weigh isolates, based on their antibiogram. The `key_antibiotics()` function adds a vector with 18 key antibiotics: 6 broad spectrum ones, 6 small spectrum for Gram negatives and 6 small spectrum for Gram positives. These can be defined by the user.

If a column exists with a name like 'key(...)ab' the `first_isolate()` function will automatically use it and determine the first weighted isolates. Mind the NOTEs in below output: 

```{r 1st weighted}
data <- data %>% 
  mutate(keyab = key_antibiotics(.)) %>% 
  mutate(first_weighted = first_isolate(.))
```

```{r, echo = FALSE, message = FALSE, warning = FALSE, results = 'asis'}
weighted_df2 <- data %>%
  filter(bacteria == as.mo("E. coli")) %>% 
  # only most prevalent patient
  filter(patient_id == top_freq(freq(., patient_id), 1)[1]) %>% 
  arrange(date) %>%
  select(date, patient_id, bacteria, amox:gent, first, first_weighted) %>% 
  # maximum of 10 rows
  .[1:min(10, nrow(.)),] %>% 
  mutate(isolate = row_number()) %>% 
  select(isolate, everything())

weighted_df2 %>% 
  knitr::kable(align = "c")
```

Instead of `r sum(weighted_df$first)`, now `r sum(weighted_df2$first_weighted)` isolates are flagged. In total, `r AMR:::percent(sum(data$first_weighted) / nrow(data))` of all isolates are marked 'first weighted' - `r AMR:::percent((sum(data$first_weighted) / nrow(data)) - (sum(data$first) / nrow(data)))` more than when using the CLSI guideline. In real life, this novel algorithm will yield 5-10% more isolates than the classic CLSI guideline.

As with `filter_first_isolate()`, there's a shortcut for this new algorithm too:
```{r 1st isolate filter 3, results = 'hide', message = FALSE, warning = FALSE}
data_1st <- data %>% 
  filter_first_weighted_isolate()
```

So we end up with `r format(nrow(data_1st), big.mark = ",")` isolates for analysis. 

We can remove unneeded columns:

```{r}
data_1st <- data_1st %>% 
  select(-c(first, keyab))
```

Now our data looks like:

```{r preview data set 3, eval = FALSE}
head(data_1st)
```

```{r preview data set 4, echo = FALSE, results = 'asis'}
knitr::kable(head(data_1st), align = "c")
```

Time for the analysis!

# Analysing the data
You might want to start by getting an idea of how the data is distributed. It's an important start, because it also decides how you will continue your analysis. 

## Dispersion of species
To just get an idea how the species are distributed, create a frequency table with our `freq()` function. We created the `genus` and `species` column earlier based on the microbial ID. With `paste()`, we can concatenate them together.

The `freq()` function can be used like the base R language was intended:

```{r freq 1, eval = FALSE}
freq(paste(data_1st$genus, data_1st$species))
```

Or can be used like the `dplyr` way, which is easier readable:

```{r freq 2a, eval = FALSE}
data_1st %>% freq(genus, species)
```
```{r freq 2b, results = 'asis', echo = FALSE}
data_1st %>% 
  freq(genus, species, header = TRUE)
```

## Resistance percentages

The functions `portion_S()`, `portion_SI()`, `portion_I()`, `portion_IR()` and `portion_R()` can be used to determine the portion of a specific antimicrobial outcome. They can be used on their own:

```{r}
data_1st %>% portion_IR(amox)
```

Or can be used in conjuction with `group_by()` and `summarise()`, both from the `dplyr` package:

```{r, eval = FALSE}
data_1st %>% 
  group_by(hospital) %>% 
  summarise(amoxicillin = portion_IR(amox))
```
```{r, echo = FALSE}
data_1st %>% 
  group_by(hospital) %>% 
  summarise(amoxicillin = portion_IR(amox)) %>% 
  knitr::kable(align = "c", big.mark = ",")
```

Of course it would be very convenient to know the number of isolates responsible for the percentages. For that purpose the `n_rsi()` can be used, which works exactly like `n_distinct()` from the `dplyr` package. It counts all isolates available for every group (i.e. values S, I or R):

```{r, eval = FALSE}
data_1st %>% 
  group_by(hospital) %>% 
  summarise(amoxicillin = portion_IR(amox),
            available = n_rsi(amox))
```
```{r, echo = FALSE}
data_1st %>% 
  group_by(hospital) %>% 
  summarise(amoxicillin = portion_IR(amox),
            available = n_rsi(amox)) %>% 
  knitr::kable(align = "c", big.mark = ",")
```

These functions can also be used to get the portion of multiple antibiotics, to calculate empiric susceptibility of combination therapies very easily:

```{r, eval = FALSE}
data_1st %>% 
  group_by(genus) %>% 
  summarise(amoxiclav = portion_S(amcl),
            gentamicin = portion_S(gent),
            amoxiclav_genta = portion_S(amcl, gent))
```
```{r, echo = FALSE}
data_1st %>% 
  group_by(genus) %>% 
  summarise(amoxiclav = portion_S(amcl),
            gentamicin = portion_S(gent),
            amoxiclav_genta = portion_S(amcl, gent)) %>% 
  knitr::kable(align = "c", big.mark = ",")
```

To make a transition to the next part, let's see how this difference could be plotted:

```{r plot 1}
data_1st %>% 
  group_by(genus) %>% 
  summarise("1. Amoxi/clav" = portion_S(amcl),
            "2. Gentamicin" = portion_S(gent),
            "3. Amoxi/clav + gent" = portion_S(amcl, gent)) %>% 
  tidyr::gather("Antibiotic", "S", -genus) %>%
  ggplot(aes(x = genus,
             y = S,
             fill = Antibiotic)) +
  geom_col(position = "dodge2")
```

## Plots
To show results in plots, most R users would nowadays use the `ggplot2` package. This package lets you create plots in layers. You can read more about it [on their website](https://ggplot2.tidyverse.org/). A quick example would look like these syntaxes:

```{r plot 2, eval = FALSE}
ggplot(data = a_data_set,
       mapping = aes(x = year,
                     y = value)) +
  geom_col() +
  labs(title = "A title",
       subtitle = "A subtitle",
       x = "My X axis",
       y = "My Y axis")

# or as short as:
ggplot(a_data_set) +
  geom_bar(aes(year))
```

The `AMR` package contains functions to extend this `ggplot2` package, for example `geom_rsi()`. It automatically transforms data with `count_df()` or `portion_df()` and show results in stacked bars. Its simplest and shortest example:

```{r plot 3}
ggplot(data_1st) +
  geom_rsi(translate_ab = FALSE)
```

Omit the `translate_ab = FALSE` to have the antibiotic codes (amox, amcl, cipr, gent) translated to official WHO names (amoxicillin, amoxicillin and betalactamase inhibitor, ciprofloxacin, gentamicin).

If we group on e.g. the `genus` column and add some additional functions from our package, we can create this:

```{r plot 4}
# group the data on `genus`
ggplot(data_1st %>% group_by(genus)) + 
  # create bars with genus on x axis
  # it looks for variables with class `rsi`,
  # of which we have 4 (earlier created with `as.rsi`)
  geom_rsi(x = "genus") + 
  # split plots on antibiotic
  facet_rsi(facet = "Antibiotic") +
  # make R red, I yellow and S green
  scale_rsi_colours() +
  # show percentages on y axis
  scale_y_percent(breaks = 0:4 * 25) +
  # turn 90 degrees, make it bars instead of columns
  coord_flip() +
  # add labels
  labs(title = "Resistance per genus and antibiotic", 
       subtitle = "(this is fake data)") +
  # and print genus in italic to follow our convention
  # (is now y axis because we turned the plot)
  theme(axis.text.y = element_text(face = "italic"))
```

To simplify this, we also created the `ggplot_rsi()` function, which combines almost all above functions:

```{r plot 5}
data_1st %>% 
  group_by(genus) %>%
  ggplot_rsi(x = "genus",
             facet = "Antibiotic",
             breaks = 0:4 * 25,
             datalabels = FALSE) +
  coord_flip()
```

## Independence test

The next example uses the included `septic_patients`, which is an anonymised data set containing 2,000 microbial blood culture isolates with their full antibiograms found in septic patients in 4 different hospitals in the Netherlands, between 2001 and 2017. It is true, genuine data. This `data.frame` can be used to practice AMR analysis.

We will compare the resistance to fosfomycin (column `fosf`) in hospital A and D. The input for the final `fisher.test()` will be this:

```{r, echo = FALSE, results = 'asis'}
septic_patients %>%
  filter(hospital_id %in% c("A", "D")) %>%
  select(hospital_id, fosf) %>%
  group_by(hospital_id) %>%
  count_df(combine_IR = TRUE) %>%
  tidyr::spread(hospital_id, Value) %>%
  select(A, D) %>%
  bind_cols(tibble(" " = c("IR", "S")), .) %>% 
  as.matrix() %>%
  knitr::kable()
```

We can transform the data and apply the test in only a couple of lines: 

```{r}
septic_patients %>%
  filter(hospital_id %in% c("A", "D")) %>% # filter on only hospitals A and D
  select(hospital_id, fosf) %>%            # select the hospitals and fosfomycin
  group_by(hospital_id) %>%                # group on the hospitals
  count_df(combine_IR = TRUE) %>%          # count all isolates per group (hospital_id)
  tidyr::spread(hospital_id, Value) %>%    # transform output so A and D are columns
  select(A, D) %>%                         # and select these only
  as.matrix() %>%                          # transform to good old matrix for fisher.test()
  fisher.test()                            # do Fisher's Exact Test
```

As can be seen, the p value is 0.03, which means that the fosfomycin resistances found in hospital A and D are really different.