---
title: "How to predict antimicrobial resistance"
author: "Matthijs S. Berends"
date: '`r format(Sys.Date(), "%d %B %Y")`'
output: 
  rmarkdown::html_vignette:
    toc: true
vignette: >
  %\VignetteIndexEntry{How to predict antimicrobial resistance}
  %\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 = 4.5
)
```

## 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"))
```

## Prediction analysis
Our package contains a function `resistance_predict()`, which takes the same input as functions for [other AMR analysis](./articles/AMR.html). Based on a date column, it calculates cases per year and uses a regression model to predict antimicrobial resistance.

It is basically as easy as:
```{r, eval = FALSE}
# resistance prediction of piperacillin/tazobactam (pita):
resistance_predict(tbl = septic_patients, col_date = "date", col_ab = "pita")

# or:
septic_patients %>% 
  resistance_predict(col_ab = "pita")

# to bind it to object 'predict_pita' for example:
predict_pita <- septic_patients %>% 
  resistance_predict(col_ab = "pita")
```

```{r, echo = FALSE}
predict_pita <- septic_patients %>% 
  resistance_predict(col_ab = "pita")
```

The function will look for a data column itself if `col_date` is not set. The result is nothing more than a `data.frame`, containing the years, number of observations, actual observed resistance, the estimated resistance and the standard error below and above the estimation:
```{r}
predict_pita
```

The function `plot` is available in base R, and can be extended by other packages to depend the output based on the type of input. We extended its function to cope with resistance predictions:

```{r}
plot(predict_pita)
```

We also support the `ggplot2` package with the function `ggplot_rsi_predict()`:

```{r}
library(ggplot2)
ggplot_rsi_predict(predict_pita)
```

### Choosing the right model

Resistance is not easily predicted; if we look at vancomycin resistance in Gram positives, the spread (i.e. standard error) is enormous:

```{r}
septic_patients %>%
  filter(mo_gramstain(mo) == "Gram positive") %>%
  resistance_predict(col_ab = "vanc", year_min = 2010, info = FALSE) %>% 
  plot()
```

Vancomycin resistance could be 100% in ten years, but might also stay around 0%. 

You can define the model with the `model` parameter. The default model is a generalised linear regression model using a binomial distribution, assuming that a period of zero resistance was followed by a period of increasing resistance leading slowly to more and more resistance.

Valid values are:

| Input values                           | Function used by R            | Type of model                                       |
|----------------------------------------|-------------------------------|-----------------------------------------------------|
| `"binomial"` or `"binom"` or `"logit"` | `glm(..., family = binomial)` | Generalised linear model with binomial distribution |
| `"loglin"` or `"poisson"`              | `glm(..., family = poisson)`  | Generalised linear model with poisson distribution  |
| `"lin"` or `"linear"`                  | `lm()`                        | Linear model                                        |

For the vancomycin resistance in Gram positive bacteria, a linear model might be more appropriate since no (left half of a) binomial distribution is to be expected based on observed years:

```{r}
septic_patients %>%
  filter(mo_gramstain(mo) == "Gram positive") %>%
  resistance_predict(col_ab = "vanc", year_min = 2010, info = FALSE, model = "linear") %>% 
  plot()
```

This seems more likely, doesn't it?